A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 20:30 UTC
[Kernel] [Helion] [15/N] Split config files into per-platform files (#36698)
Signed-off-by: Yanan Cao <gmagogsfm@gmail.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Yanan Cao committed
Mar 11, 2026 at 14:25 UTC
cf632499ee31e50f421fe21127876688290c6496
5 files changed
+27833
-27804
tests/kernels/helion/test_config_manager.py
+38
-33
@@ -160,10 +160,11 @@ class TestConfigManager:
160
"""Test getting config file path for a kernel."""
161
manager = ConfigManager(base_dir="/tmp")
162
163
- file_path = manager.get_config_file_path("silu_mul_fp8")
163
+ dir_path = manager.get_config_file_path("silu_mul_fp8")
164
+ assert dir_path == Path("/tmp/silu_mul_fp8")
165
165
- expected_path = Path("/tmp/silu_mul_fp8.json")
166
- assert file_path == expected_path
166
+ file_path = manager.get_config_file_path("silu_mul_fp8", "nvidia_h100")
167
+ assert file_path == Path("/tmp/silu_mul_fp8/nvidia_h100.json")
168
169
def test_ensure_base_dir_exists(self):
170
"""Test ensuring base directory exists."""
@@ -189,19 +190,19 @@ class TestConfigManager:
190
assert config_set.get_platforms() == []
191
192
def test_load_config_set_valid_file(self):
192
- """Test loading config set from valid file."""
193
+ """Test loading config set from per-platform files."""
194
with tempfile.TemporaryDirectory() as temp_dir:
194
- # Use realistic config data
195
kernel_config = {
196
"block_sizes": [128, 64],
197
"num_warps": 8,
198
"num_stages": 6,
199
"pid_type": "persistent_interleaved",
200
}
201
- config_data = {"h100": {"batch_32_hidden_4096": kernel_config}}
202
- config_file = Path(temp_dir) / "test_kernel.json"
203
- with open(config_file, "w") as f:
204
- json.dump(config_data, f)
201
+ kernel_dir = Path(temp_dir) / "test_kernel"
202
+ kernel_dir.mkdir()
203
+ platform_file = kernel_dir / "h100.json"
204
+ with open(platform_file, "w") as f:
205
+ json.dump({"batch_32_hidden_4096": kernel_config}, f)
206
207
manager = ConfigManager(base_dir=temp_dir)
208
config_set = manager.load_config_set("test_kernel")
@@ -210,7 +211,6 @@ class TestConfigManager:
211
assert config_set.kernel_name == "test_kernel"
212
assert config_set.get_platforms() == ["h100"]
213
213
- # Verify the config was loaded correctly
214
config = config_set.get_config("h100", "batch_32_hidden_4096")
215
assert isinstance(config, helion.Config)
216
assert config.block_sizes == [128, 64]
@@ -219,7 +219,9 @@ class TestConfigManager:
219
def test_load_config_set_invalid_json(self):
220
"""Test loading config set from file with invalid JSON."""
221
with tempfile.TemporaryDirectory() as temp_dir:
222
- config_file = Path(temp_dir) / "test_kernel.json"
222
+ kernel_dir = Path(temp_dir) / "test_kernel"
223
+ kernel_dir.mkdir()
224
+ config_file = kernel_dir / "h100.json"
225
with open(config_file, "w") as f:
226
f.write("invalid json content {")
227
@@ -231,9 +233,8 @@ class TestConfigManager:
233
assert config_set.get_platforms() == []
234
235
def test_save_config_set(self):
234
- """Test saving ConfigSet to file."""
236
+ """Test saving ConfigSet to per-platform files."""
237
with tempfile.TemporaryDirectory() as temp_dir:
236
- # Use realistic config data
238
kernel_config = {
239
"block_sizes": [256, 128],
240
"num_warps": 16,
@@ -246,31 +247,34 @@ class TestConfigManager:
247
manager = ConfigManager(base_dir=temp_dir)
248
saved_path = manager.save_config_set(config_set)
249
249
- expected_path = Path(temp_dir) / "test_kernel.json"
250
- assert saved_path == expected_path
251
- assert saved_path.exists()
250
+ expected_dir = Path(temp_dir) / "test_kernel"
251
+ assert saved_path == expected_dir
252
+ assert saved_path.is_dir()
253
253
- with open(saved_path) as f:
254
+ platform_file = expected_dir / "h100.json"
255
+ assert platform_file.exists()
256
+ with open(platform_file) as f:
257
loaded_data = json.load(f)
255
- assert loaded_data == data
258
+ assert loaded_data == data["h100"]
259
260
def test_save_config_set_creates_directory(self):
261
"""Test that save_config_set creates parent directories if needed."""
262
with tempfile.TemporaryDirectory() as temp_dir:
263
nested_dir = Path(temp_dir) / "nested" / "configs"
261
- config_set = ConfigSet("test_kernel")
264
+ data = {"h100": {"default": {"num_warps": 4}}}
265
+ config_set = ConfigSet.from_dict("test_kernel", data)
266
267
manager = ConfigManager(base_dir=nested_dir)
268
saved_path = manager.save_config_set(config_set)
269
270
assert nested_dir.exists()
271
assert nested_dir.is_dir()
268
- assert saved_path.exists()
272
+ assert saved_path.is_dir()
273
+ assert (saved_path / "h100.json").exists()
274
275
def test_get_platform_configs(self):
276
"""Test getting all configs for a specific platform."""
277
with tempfile.TemporaryDirectory() as temp_dir:
273
- # Use realistic config data
278
config_1 = {"num_warps": 4, "num_stages": 3, "block_sizes": [64, 32]}
279
config_2 = {"num_warps": 8, "num_stages": 5, "block_sizes": [128, 64]}
280
default_config = {
@@ -280,17 +284,19 @@ class TestConfigManager:
284
}
285
config_3 = {"num_warps": 2, "num_stages": 2, "block_sizes": [32, 16]}
286
283
- config_data = {
284
- "h100": {
285
- "batch_32_hidden_4096": config_1,
286
- "batch_64_hidden_2048": config_2,
287
- "default": default_config,
288
- },
289
- "a100": {"batch_16_hidden_1024": config_3},
290
- }
291
- config_file = Path(temp_dir) / "test_kernel.json"
292
- with open(config_file, "w") as f:
293
- json.dump(config_data, f)
287
+ kernel_dir = Path(temp_dir) / "test_kernel"
288
+ kernel_dir.mkdir()
289
+ with open(kernel_dir / "h100.json", "w") as f:
290
+ json.dump(
291
+ {
292
+ "batch_32_hidden_4096": config_1,
293
+ "batch_64_hidden_2048": config_2,
294
+ "default": default_config,
295
+ },
296
+ f,
297
+ )
298
+ with open(kernel_dir / "a100.json", "w") as f:
299
+ json.dump({"batch_16_hidden_1024": config_3}, f)
300
301
manager = ConfigManager(base_dir=temp_dir)
302
@@ -302,7 +308,6 @@ class TestConfigManager:
308
for config in h100_configs.values():
309
assert isinstance(config, helion.Config)
310
305
- # Verify specific config details
311
assert h100_configs["batch_32_hidden_4096"].num_warps == 4
312
assert h100_configs["default"].num_stages == 7
313
vllm/kernels/helion/config_manager.py
+63
-37
@@ -8,23 +8,15 @@ operations, including naming conventions, directory resolution, and file I/O.
8
9
Config File Structure
10
---------------------
11
-Each kernel has a single JSON config file: {kernel_name}.json
12
-
13
-The file uses a simplified 2-layer hierarchical structure:
14
-{
15
- "h100": { # GPU platform
16
- "default": { ... }, # Fallback configuration
17
- "batch_32_hidden_4096": { ... },
18
- "batch_64_hidden_8192": { ... }
19
- },
20
- "a100": {
21
- "default": { ... },
22
- "batch_16_hidden_2048": { ... }
23
- }
24
-}
25
-
26
-Example file: silu_mul_fp8.json
11
+Each kernel has a directory: {kernel_name}/
12
+Inside, each GPU platform has its own JSON file: {kernel_name}/{platform}.json
13
14
+For example:
15
+ silu_mul_fp8/
16
+ nvidia_h100.json # { "default": {...}, "batch_32_hidden_4096": {...} }
17
+ nvidia_h200.json # { "batch_16_hidden_2048": {...} }
18
+
19
+Each platform file maps config keys to Helion config objects.
20
Config keys should be structured strings that encode the relevant
21
parameters (e.g., "batch_32_hidden_4096", "seq_512_heads_16", "fp8_batch_64", etc.).
22
@@ -212,8 +204,15 @@ class ConfigManager:
204
cls._instance = None
205
cls._instance_base_dir = None
206
215
- def get_config_file_path(self, kernel_name: str) -> Path:
216
- return self._base_dir / f"{kernel_name}.json"
207
+ def get_kernel_dir(self, kernel_name: str) -> Path:
208
+ return self._base_dir / kernel_name
209
+
210
+ def get_config_file_path(
211
+ self, kernel_name: str, platform: str | None = None
212
+ ) -> Path:
213
+ if platform is not None:
214
+ return self.get_kernel_dir(kernel_name) / f"{platform}.json"
215
+ return self.get_kernel_dir(kernel_name)
216
217
def ensure_base_dir_exists(self) -> Path:
218
self._base_dir.mkdir(parents=True, exist_ok=True)
@@ -230,39 +229,59 @@ class ConfigManager:
229
f"Config directory '{self._base_dir}' is not writable: {e}"
230
) from e
231
233
- def load_config_set(self, kernel_name: str) -> ConfigSet:
234
- config_path = self.get_config_file_path(kernel_name)
232
+ def _load_platform_file(self, kernel_name: str, platform: str) -> dict[str, Any]:
233
+ config_path = self.get_config_file_path(kernel_name, platform)
234
if not config_path.exists():
236
- return ConfigSet.from_dict(kernel_name, {})
237
-
235
+ return {}
236
try:
237
with open(config_path) as f:
240
- data = json.load(f)
241
- return ConfigSet.from_dict(kernel_name, data)
238
+ return json.load(f)
239
except (json.JSONDecodeError, OSError) as e:
240
logger.error("Failed to load config file %s: %s", config_path, e)
241
+ return {}
242
+
243
+ def load_config_set(self, kernel_name: str) -> ConfigSet:
244
+ kernel_dir = self.get_kernel_dir(kernel_name)
245
+ if not kernel_dir.is_dir():
246
return ConfigSet.from_dict(kernel_name, {})
247
248
+ data: dict[str, Any] = {}
249
+ for platform_file in sorted(kernel_dir.glob("*.json")):
250
+ platform = platform_file.stem
251
+ try:
252
+ with open(platform_file) as f:
253
+ platform_data = json.load(f)
254
+ data[platform] = platform_data
255
+ except (json.JSONDecodeError, OSError) as e:
256
+ logger.error("Failed to load config file %s: %s", platform_file, e)
257
+
258
+ return ConfigSet.from_dict(kernel_name, data)
259
+
260
def get_platform_configs(
261
self, kernel_name: str, platform: str
262
) -> dict[str, helion.Config]:
249
- config_set = self.load_config_set(kernel_name)
263
+ platform_data = self._load_platform_file(kernel_name, platform)
264
+ if not platform_data:
265
+ return {}
266
+ config_set = ConfigSet.from_dict(kernel_name, {platform: platform_data})
267
config_keys = config_set.get_config_keys(platform)
251
-
268
return {
269
config_key: config_set.get_config(platform, config_key)
270
for config_key in config_keys
271
}
272
273
def save_config_set(self, config_set: ConfigSet) -> Path:
258
- config_path = self.get_config_file_path(config_set.kernel_name)
259
- config_path.parent.mkdir(parents=True, exist_ok=True)
274
+ kernel_dir = self.get_kernel_dir(config_set.kernel_name)
275
+ kernel_dir.mkdir(parents=True, exist_ok=True)
276
261
- with open(config_path, "w") as f:
262
- json.dump(config_set.to_dict(), f, indent=2)
277
+ full_data = config_set.to_dict()
278
+ for platform, platform_data in full_data.items():
279
+ platform_path = kernel_dir / f"{platform}.json"
280
+ with open(platform_path, "w") as f:
281
+ json.dump(platform_data, f, indent=2)
282
+ logger.info("Saved config to: %s", platform_path)
283
264
- logger.info("Saved config to: %s", config_path)
265
- return config_path
284
+ return kernel_dir
285
286
def save_configs(
287
self,
@@ -271,11 +290,18 @@ class ConfigManager:
290
configs: dict[str, "helion.Config"],
291
) -> Path:
292
"""Save configs for a kernel/platform, merging with existing."""
274
- config_set = self.load_config_set(kernel_name)
293
+ platform_data = self._load_platform_file(kernel_name, platform)
294
for config_key, config in configs.items():
276
- config_set.set_config(platform, config_key, config)
277
- return self.save_config_set(config_set)
295
+ platform_data[config_key] = json.loads(config.to_json())
296
+
297
+ platform_path = self.get_config_file_path(kernel_name, platform)
298
+ platform_path.parent.mkdir(parents=True, exist_ok=True)
299
+ with open(platform_path, "w") as f:
300
+ json.dump(platform_data, f, indent=2)
301
+
302
+ logger.info("Saved config to: %s", platform_path)
303
+ return platform_path
304
305
def config_exists(self, kernel_name: str, platform: str, config_key: str) -> bool:
280
- config_set = self.load_config_set(kernel_name)
281
- return config_set.has_config(platform, config_key)
306
+ platform_data = self._load_platform_file(kernel_name, platform)
307
+ return config_key in platform_data
vllm/kernels/helion/configs/silu_mul_fp8.json
deleted
-27734
@@ -1,27734 +0,0 @@
1
-{
2
- "nvidia_h200": {
3
- "intermediate_2048_numtokens_256": {
4
- "block_sizes": [
5
- 64,
6
- 32
7
- ],
8
- "loop_orders": [
9
- [
10
- 0,
11
- 1
12
- ]
13
- ],
14
- "flatten_loops": [
15
- true
16
- ],
17
- "l2_groupings": [
18
- 1
19
- ],
20
- "range_unroll_factors": [
21
- 0
22
- ],
23
- "range_warp_specializes": [],
24
- "range_num_stages": [
25
- 0
26
- ],
27
- "range_multi_buffers": [
28
- null
29
- ],
30
- "range_flattens": [
31
- null
32
- ],
33
- "load_eviction_policies": [
34
- "",
35
- "",
36
- ""
37
- ],
38
- "num_warps": 8,
39
- "num_stages": 1,
40
- "indexing": [
41
- "pointer",
42
- "pointer",
43
- "pointer",
44
- "pointer"
45
- ],
46
- "pid_type": "flat"
47
- },
48
- "intermediate_4096_numtokens_256": {
49
- "block_sizes": [
50
- 256,
51
- 32
52
- ],
53
- "loop_orders": [
54
- [
55
- 0,
56
- 1
57
- ]
58
- ],
59
- "flatten_loops": [
60
- true
61
- ],
62
- "l2_groupings": [
63
- 1
64
- ],
65
- "range_unroll_factors": [
66
- 0
67
- ],
68
- "range_warp_specializes": [],
69
- "range_num_stages": [
70
- 0
71
- ],
72
- "range_multi_buffers": [
73
- null
74
- ],
75
- "range_flattens": [
76
- null
77
- ],
78
- "load_eviction_policies": [
79
- "",
80
- "",
81
- ""
82
- ],
83
- "num_warps": 8,
84
- "num_stages": 1,
85
- "indexing": [
86
- "pointer",
87
- "pointer",
88
- "pointer",
89
- "pointer"
90
- ],
91
- "pid_type": "flat"
92
- },
93
- "default": {
94
- "block_sizes": [
95
- 1,
96
- 512
97
- ],
98
- "loop_orders": [
99
- [
100
- 1,
101
- 0
102
- ]
103
- ],
104
- "flatten_loops": [
105
- false
106
- ],
107
- "l2_groupings": [
108
- 4
109
- ],
110
- "range_unroll_factors": [
111
- 0
112
- ],
113
- "range_warp_specializes": [],
114
- "range_num_stages": [
115
- 0
116
- ],
117
- "range_multi_buffers": [
118
- null
119
- ],
120
- "range_flattens": [
121
- null
122
- ],
123
- "load_eviction_policies": [
124
- "",
125
- "first",
126
- ""
127
- ],
128
- "num_warps": 8,
129
- "num_stages": 2,
130
- "indexing": [
131
- "tensor_descriptor",
132
- "tensor_descriptor",
133
- "tensor_descriptor",
134
- "pointer"
135
- ],
136
- "pid_type": "flat"
137
- },
138
- "intermediate_8192_numtokens_256": {
139
- "block_sizes": [
140
- 256,
141
- 64
142
- ],
143
- "loop_orders": [
144
- [
145
- 0,
146
- 1
147
- ]
148
- ],
149
- "flatten_loops": [
150
- true
151
- ],
152
- "l2_groupings": [
153
- 1
154
- ],
155
- "range_unroll_factors": [
156
- 0
157
- ],
158
- "range_warp_specializes": [],
159
- "range_num_stages": [
160
- 0
161
- ],
162
- "range_multi_buffers": [
163
- null
164
- ],
165
- "range_flattens": [
166
- null
167
- ],
168
- "load_eviction_policies": [
169
- "",
170
- "",
171
- ""
172
- ],
173
- "num_warps": 8,
174
- "num_stages": 1,
175
- "indexing": [
176
- "pointer",
177
- "pointer",
178
- "pointer",
179
- "pointer"
180
- ],
181
- "pid_type": "flat"
182
- },
183
- "intermediate_14336_numtokens_256": {
184
- "block_sizes": [
185
- 8,
186
- 4096
187
- ],
188
- "loop_orders": [
189
- [
190
- 0,
191
- 1
192
- ]
193
- ],
194
- "flatten_loops": [
195
- true
196
- ],
197
- "l2_groupings": [
198
- 1
199
- ],
200
- "range_unroll_factors": [
201
- 0
202
- ],
203
- "range_warp_specializes": [],
204
- "range_num_stages": [
205
- 0
206
- ],
207
- "range_multi_buffers": [
208
- null
209
- ],
210
- "range_flattens": [
211
- null
212
- ],
213
- "load_eviction_policies": [
214
- "",
215
- "",
216
- ""
217
- ],
218
- "num_warps": 8,
219
- "num_stages": 1,
220
- "indexing": [
221
- "pointer",
222
- "pointer",
223
- "pointer",
224
- "pointer"
225
- ],
226
- "pid_type": "flat"
227
- },
228
- "intermediate_11008_numtokens_256": {
229
- "block_sizes": [
230
- 64,
231
- 32
232
- ],
233
- "loop_orders": [
234
- [
235
- 0,
236
- 1
237
- ]
238
- ],
239
- "flatten_loops": [
240
- true
241
- ],
242
- "l2_groupings": [
243
- 1
244
- ],
245
- "range_unroll_factors": [
246
- 0
247
- ],
248
- "range_warp_specializes": [],
249
- "range_num_stages": [
250
- 0
251
- ],
252
- "range_multi_buffers": [
253
- null
254
- ],
255
- "range_flattens": [
256
- null
257
- ],
258
- "load_eviction_policies": [
259
- "",
260
- "",
261
- ""
262
- ],
263
- "num_warps": 2,
264
- "num_stages": 1,
265
- "indexing": [
266
- "pointer",
267
- "pointer",
268
- "pointer",
269
- "pointer"
270
- ],
271
- "pid_type": "flat"
272
- },
273
- "intermediate_7688_numtokens_256": {
274
- "block_sizes": [
275
- 32,
276
- 512
277
- ],
278
- "loop_orders": [
279
- [
280
- 0,
281
- 1
282
- ]
283
- ],
284
- "flatten_loops": [
285
- true
286
- ],
287
- "l2_groupings": [
288
- 1
289
- ],
290
- "range_unroll_factors": [
291
- 0
292
- ],
293
- "range_warp_specializes": [],
294
- "range_num_stages": [
295
- 0
296
- ],
297
- "range_multi_buffers": [
298
- null
299
- ],
300
- "range_flattens": [
301
- null
302
- ],
303
- "load_eviction_policies": [
304
- "",
305
- "",
306
- ""
307
- ],
308
- "num_warps": 8,
309
- "num_stages": 1,
310
- "indexing": [
311
- "pointer",
312
- "pointer",
313
- "pointer",
314
- "pointer"
315
- ],
316
- "pid_type": "flat"
317
- },
318
- "intermediate_2880_numtokens_256": {
319
- "block_sizes": [
320
- 32,
321
- 64
322
- ],
323
- "loop_orders": [
324
- [
325
- 0,
326
- 1
327
- ]
328
- ],
329
- "flatten_loops": [
330
- true
331
- ],
332
- "l2_groupings": [
333
- 2
334
- ],
335
- "range_unroll_factors": [
336
- 0
337
- ],
338
- "range_warp_specializes": [],
339
- "range_num_stages": [
340
- 0
341
- ],
342
- "range_multi_buffers": [
343
- null
344
- ],
345
- "range_flattens": [
346
- null
347
- ],
348
- "load_eviction_policies": [
349
- "",
350
- "",
351
- ""
352
- ],
353
- "num_warps": 8,
354
- "num_stages": 1,
355
- "indexing": [
356
- "pointer",
357
- "pointer",
358
- "pointer",
359
- "pointer"
360
- ],
361
- "pid_type": "flat"
362
- },
363
- "intermediate_2048_numtokens_1": {
364
- "block_sizes": [
365
- 1,
366
- 256
367
- ],
368
- "loop_orders": [
369
- [
370
- 0,
371
- 1
372
- ]
373
- ],
374
- "flatten_loops": [
375
- true
376
- ],
377
- "l2_groupings": [
378
- 1
379
- ],
380
- "range_unroll_factors": [
381
- 0
382
- ],
383
- "range_warp_specializes": [],
384
- "range_num_stages": [
385
- 0
386
- ],
387
- "range_multi_buffers": [
388
- null
389
- ],
390
- "range_flattens": [
391
- null
392
- ],
393
- "load_eviction_policies": [
394
- "",
395
- "",
396
- ""
397
- ],
398
- "num_warps": 2,
399
- "num_stages": 1,
400
- "indexing": [
401
- "pointer",
402
- "pointer",
403
- "pointer",
404
- "pointer"
405
- ],
406
- "pid_type": "flat"
407
- },
408
- "intermediate_2880_numtokens_1": {
409
- "block_sizes": [
410
- 1,
411
- 128
412
- ],
413
- "loop_orders": [
414
- [
415
- 0,
416
- 1
417
- ]
418
- ],
419
- "flatten_loops": [
420
- true
421
- ],
422
- "l2_groupings": [
423
- 1
424
- ],
425
- "range_unroll_factors": [
426
- 0
427
- ],
428
- "range_warp_specializes": [],
429
- "range_num_stages": [
430
- 0
431
- ],
432
- "range_multi_buffers": [
433
- null
434
- ],
435
- "range_flattens": [
436
- null
437
- ],
438
- "load_eviction_policies": [
439
- "",
440
- "",
441
- ""
442
- ],
443
- "num_warps": 4,
444
- "num_stages": 1,
445
- "indexing": [
446
- "pointer",
447
- "pointer",
448
- "pointer",
449
- "pointer"
450
- ],
451
- "pid_type": "flat"
452
- },
453
- "intermediate_4096_numtokens_1": {
454
- "block_sizes": [
455
- 1,
456
- 256
457
- ],
458
- "loop_orders": [
459
- [
460
- 0,
461
- 1
462
- ]
463
- ],
464
- "flatten_loops": [
465
- true
466
- ],
467
- "l2_groupings": [
468
- 1
469
- ],
470
- "range_unroll_factors": [
471
- 0
472
- ],
473
- "range_warp_specializes": [],
474
- "range_num_stages": [
475
- 0
476
- ],
477
- "range_multi_buffers": [
478
- null
479
- ],
480
- "range_flattens": [
481
- null
482
- ],
483
- "load_eviction_policies": [
484
- "",
485
- "",
486
- ""
487
- ],
488
- "num_warps": 2,
489
- "num_stages": 1,
490
- "indexing": [
491
- "pointer",
492
- "pointer",
493
- "pointer",
494
- "pointer"
495
- ],
496
- "pid_type": "flat"
497
- },
498
- "intermediate_8192_numtokens_1": {
499
- "block_sizes": [
500
- 1,
501
- 256
502
- ],
503
- "loop_orders": [
504
- [
505
- 0,
506
- 1
507
- ]
508
- ],
509
- "flatten_loops": [
510
- true
511
- ],
512
- "l2_groupings": [
513
- 1
514
- ],
515
- "range_unroll_factors": [
516
- 0
517
- ],
518
- "range_warp_specializes": [],
519
- "range_num_stages": [
520
- 0
521
- ],
522
- "range_multi_buffers": [
523
- null
524
- ],
525
- "range_flattens": [
526
- null
527
- ],
528
- "load_eviction_policies": [
529
- "",
530
- "",
531
- ""
532
- ],
533
- "num_warps": 2,
534
- "num_stages": 1,
535
- "indexing": [
536
- "pointer",
537
- "pointer",
538
- "pointer",
539
- "pointer"
540
- ],
541
- "pid_type": "flat"
542
- },
543
- "intermediate_11008_numtokens_1": {
544
- "block_sizes": [
545
- 1,
546
- 256
547
- ],
548
- "loop_orders": [
549
- [
550
- 0,
551
- 1
552
- ]
553
- ],
554
- "flatten_loops": [
555
- true
556
- ],
557
- "l2_groupings": [
558
- 1
559
- ],
560
- "range_unroll_factors": [
561
- 0
562
- ],
563
- "range_warp_specializes": [],
564
- "range_num_stages": [
565
- 0
566
- ],
567
- "range_multi_buffers": [
568
- null
569
- ],
570
- "range_flattens": [
571
- null
572
- ],
573
- "load_eviction_policies": [
574
- "",
575
- "",
576
- ""
577
- ],
578
- "num_warps": 2,
579
- "num_stages": 1,
580
- "indexing": [
581
- "pointer",
582
- "pointer",
583
- "pointer",
584
- "pointer"
585
- ],
586
- "pid_type": "flat"
587
- },
588
- "intermediate_14336_numtokens_1": {
589
- "block_sizes": [
590
- 1,
591
- 256
592
- ],
593
- "loop_orders": [
594
- [
595
- 0,
596
- 1
597
- ]
598
- ],
599
- "flatten_loops": [
600
- true
601
- ],
602
- "l2_groupings": [
603
- 1
604
- ],
605
- "range_unroll_factors": [
606
- 0
607
- ],
608
- "range_warp_specializes": [],
609
- "range_num_stages": [
610
- 0
611
- ],
612
- "range_multi_buffers": [
613
- null
614
- ],
615
- "range_flattens": [
616
- null
617
- ],
618
- "load_eviction_policies": [
619
- "",
620
- "",
621
- ""
622
- ],
623
- "num_warps": 2,
624
- "num_stages": 1,
625
- "indexing": [
626
- "pointer",
627
- "pointer",
628
- "pointer",
629
- "pointer"
630
- ],
631
- "pid_type": "flat"
632
- },
633
- "intermediate_2048_numtokens_2": {
634
- "block_sizes": [
635
- 2,
636
- 128
637
- ],
638
- "loop_orders": [
639
- [
640
- 0,
641
- 1
642
- ]
643
- ],
644
- "flatten_loops": [
645
- true
646
- ],
647
- "l2_groupings": [
648
- 1
649
- ],
650
- "range_unroll_factors": [
651
- 0
652
- ],
653
- "range_warp_specializes": [],
654
- "range_num_stages": [
655
- 0
656
- ],
657
- "range_multi_buffers": [
658
- null
659
- ],
660
- "range_flattens": [
661
- null
662
- ],
663
- "load_eviction_policies": [
664
- "",
665
- "",
666
- ""
667
- ],
668
- "num_warps": 2,
669
- "num_stages": 1,
670
- "indexing": [
671
- "pointer",
672
- "pointer",
673
- "pointer",
674
- "pointer"
675
- ],
676
- "pid_type": "flat"
677
- },
678
- "intermediate_2880_numtokens_2": {
679
- "block_sizes": [
680
- 2,
681
- 256
682
- ],
683
- "loop_orders": [
684
- [
685
- 0,
686
- 1
687
- ]
688
- ],
689
- "flatten_loops": [
690
- true
691
- ],
692
- "l2_groupings": [
693
- 1
694
- ],
695
- "range_unroll_factors": [
696
- 0
697
- ],
698
- "range_warp_specializes": [],
699
- "range_num_stages": [
700
- 0
701
- ],
702
- "range_multi_buffers": [
703
- null
704
- ],
705
- "range_flattens": [
706
- null
707
- ],
708
- "load_eviction_policies": [
709
- "",
710
- "",
711
- ""
712
- ],
713
- "num_warps": 4,
714
- "num_stages": 1,
715
- "indexing": [
716
- "pointer",
717
- "pointer",
718
- "pointer",
719
- "pointer"
720
- ],
721
- "pid_type": "flat"
722
- },
723
- "intermediate_4096_numtokens_2": {
724
- "block_sizes": [
725
- 2,
726
- 128
727
- ],
728
- "loop_orders": [
729
- [
730
- 0,
731
- 1
732
- ]
733
- ],
734
- "flatten_loops": [
735
- true
736
- ],
737
- "l2_groupings": [
738
- 1
739
- ],
740
- "range_unroll_factors": [
741
- 0
742
- ],
743
- "range_warp_specializes": [],
744
- "range_num_stages": [
745
- 0
746
- ],
747
- "range_multi_buffers": [
748
- null
749
- ],
750
- "range_flattens": [
751
- null
752
- ],
753
- "load_eviction_policies": [
754
- "",
755
- "",
756
- ""
757
- ],
758
- "num_warps": 2,
759
- "num_stages": 1,
760
- "indexing": [
761
- "pointer",
762
- "pointer",
763
- "pointer",
764
- "pointer"
765
- ],
766
- "pid_type": "flat"
767
- },
768
- "intermediate_8192_numtokens_2": {
769
- "block_sizes": [
770
- 2,
771
- 128
772
- ],
773
- "loop_orders": [
774
- [
775
- 0,
776
- 1
777
- ]
778
- ],
779
- "flatten_loops": [
780
- true
781
- ],
782
- "l2_groupings": [
783
- 1
784
- ],
785
- "range_unroll_factors": [
786
- 0
787
- ],
788
- "range_warp_specializes": [],
789
- "range_num_stages": [
790
- 0
791
- ],
792
- "range_multi_buffers": [
793
- null
794
- ],
795
- "range_flattens": [
796
- null
797
- ],
798
- "load_eviction_policies": [
799
- "",
800
- "",
801
- ""
802
- ],
803
- "num_warps": 2,
804
- "num_stages": 1,
805
- "indexing": [
806
- "pointer",
807
- "pointer",
808
- "pointer",
809
- "pointer"
810
- ],
811
- "pid_type": "flat"
812
- },
813
- "intermediate_11008_numtokens_2": {
814
- "block_sizes": [
815
- 1,
816
- 16384
817
- ],
818
- "loop_orders": [
819
- [
820
- 1,
821
- 0
822
- ]
823
- ],
824
- "flatten_loops": [
825
- true
826
- ],
827
- "l2_groupings": [
828
- 16
829
- ],
830
- "range_unroll_factors": [
831
- 0
832
- ],
833
- "range_warp_specializes": [],
834
- "range_num_stages": [
835
- 0
836
- ],
837
- "range_multi_buffers": [
838
- null
839
- ],
840
- "range_flattens": [
841
- null
842
- ],
843
- "load_eviction_policies": [
844
- "last",
845
- "first",
846
- "last"
847
- ],
848
- "num_warps": 32,
849
- "num_stages": 3,
850
- "indexing": [
851
- "tensor_descriptor",
852
- "pointer",
853
- "pointer",
854
- "tensor_descriptor"
855
- ],
856
- "pid_type": "xyz"
857
- },
858
- "intermediate_14336_numtokens_2": {
859
- "block_sizes": [
860
- 2,
861
- 128
862
- ],
863
- "loop_orders": [
864
- [
865
- 0,
866
- 1
867
- ]
868
- ],
869
- "flatten_loops": [
870
- true
871
- ],
872
- "l2_groupings": [
873
- 1
874
- ],
875
- "range_unroll_factors": [
876
- 0
877
- ],
878
- "range_warp_specializes": [],
879
- "range_num_stages": [
880
- 0
881
- ],
882
- "range_multi_buffers": [
883
- null
884
- ],
885
- "range_flattens": [
886
- null
887
- ],
888
- "load_eviction_policies": [
889
- "",
890
- "",
891
- ""
892
- ],
893
- "num_warps": 2,
894
- "num_stages": 1,
895
- "indexing": [
896
- "pointer",
897
- "pointer",
898
- "pointer",
899
- "pointer"
900
- ],
901
- "pid_type": "flat"
902
- },
903
- "intermediate_2048_numtokens_4": {
904
- "block_sizes": [
905
- 4,
906
- 64
907
- ],
908
- "loop_orders": [
909
- [
910
- 0,
911
- 1
912
- ]
913
- ],
914
- "flatten_loops": [
915
- true
916
- ],
917
- "l2_groupings": [
918
- 1
919
- ],
920
- "range_unroll_factors": [
921
- 0
922
- ],
923
- "range_warp_specializes": [],
924
- "range_num_stages": [
925
- 0
926
- ],
927
- "range_multi_buffers": [
928
- null
929
- ],
930
- "range_flattens": [
931
- null
932
- ],
933
- "load_eviction_policies": [
934
- "",
935
- "",
936
- ""
937
- ],
938
- "num_warps": 2,
939
- "num_stages": 1,
940
- "indexing": [
941
- "pointer",
942
- "pointer",
943
- "pointer",
944
- "pointer"
945
- ],
946
- "pid_type": "flat"
947
- },
948
- "intermediate_2880_numtokens_4": {
949
- "block_sizes": [
950
- 4,
951
- 64
952
- ],
953
- "loop_orders": [
954
- [
955
- 0,
956
- 1
957
- ]
958
- ],
959
- "flatten_loops": [
960
- true
961
- ],
962
- "l2_groupings": [
963
- 64
964
- ],
965
- "range_unroll_factors": [
966
- 0
967
- ],
968
- "range_warp_specializes": [],
969
- "range_num_stages": [
970
- 0
971
- ],
972
- "range_multi_buffers": [
973
- null
974
- ],
975
- "range_flattens": [
976
- null
977
- ],
978
- "load_eviction_policies": [
979
- "",
980
- "last",
981
- "first"
982
- ],
983
- "num_warps": 8,
984
- "num_stages": 7,
985
- "indexing": [
986
- "pointer",
987
- "tensor_descriptor",
988
- "tensor_descriptor",
989
- "pointer"
990
- ],
991
- "pid_type": "xyz"
992
- },
993
- "intermediate_4096_numtokens_4": {
994
- "block_sizes": [
995
- 4,
996
- 64
997
- ],
998
- "loop_orders": [
999
- [
1000
- 0,
1001
- 1
1002
- ]
1003
- ],
1004
- "flatten_loops": [
1005
- true
1006
- ],
1007
- "l2_groupings": [
1008
- 1
1009
- ],
1010
- "range_unroll_factors": [
1011
- 0
1012
- ],
1013
- "range_warp_specializes": [],
1014
- "range_num_stages": [
1015
- 0
1016
- ],
1017
- "range_multi_buffers": [
1018
- null
1019
- ],
1020
- "range_flattens": [
1021
- null
1022
- ],
1023
- "load_eviction_policies": [
1024
- "",
1025
- "",
1026
- ""
1027
- ],
1028
- "num_warps": 2,
1029
- "num_stages": 1,
1030
- "indexing": [
1031
- "pointer",
1032
- "pointer",
1033
- "pointer",
1034
- "pointer"
1035
- ],
1036
- "pid_type": "flat"
1037
- },
1038
- "intermediate_8192_numtokens_4": {
1039
- "block_sizes": [
1040
- 1,
1041
- 2048
1042
- ],
1043
- "loop_orders": [
1044
- [
1045
- 0,
1046
- 1
1047
- ]
1048
- ],
1049
- "flatten_loops": [
1050
- false
1051
- ],
1052
- "l2_groupings": [
1053
- 1
1054
- ],
1055
- "range_unroll_factors": [
1056
- 0
1057
- ],
1058
- "range_warp_specializes": [],
1059
- "range_num_stages": [
1060
- 0
1061
- ],
1062
- "range_multi_buffers": [
1063
- null
1064
- ],
1065
- "range_flattens": [
1066
- null
1067
- ],
1068
- "load_eviction_policies": [
1069
- "",
1070
- "",
1071
- ""
1072
- ],
1073
- "num_warps": 8,
1074
- "num_stages": 8,
1075
- "indexing": [
1076
- "pointer",
1077
- "tensor_descriptor",
1078
- "pointer",
1079
- "pointer"
1080
- ],
1081
- "pid_type": "flat"
1082
- },
1083
- "intermediate_11008_numtokens_4": {
1084
- "block_sizes": [
1085
- 4,
1086
- 2048
1087
- ],
1088
- "loop_orders": [
1089
- [
1090
- 1,
1091
- 0
1092
- ]
1093
- ],
1094
- "flatten_loops": [
1095
- true
1096
- ],
1097
- "l2_groupings": [
1098
- 4
1099
- ],
1100
- "range_unroll_factors": [
1101
- 0
1102
- ],
1103
- "range_warp_specializes": [],
1104
- "range_num_stages": [
1105
- 0
1106
- ],
1107
- "range_multi_buffers": [
1108
- null
1109
- ],
1110
- "range_flattens": [
1111
- null
1112
- ],
1113
- "load_eviction_policies": [
1114
- "first",
1115
- "last",
1116
- "last"
1117
- ],
1118
- "num_warps": 8,
1119
- "num_stages": 6,
1120
- "indexing": [
1121
- "pointer",
1122
- "tensor_descriptor",
1123
- "tensor_descriptor",
1124
- "tensor_descriptor"
1125
- ],
1126
- "pid_type": "xyz"
1127
- },
1128
- "intermediate_14336_numtokens_4": {
1129
- "block_sizes": [
1130
- 4,
1131
- 256
1132
- ],
1133
- "loop_orders": [
1134
- [
1135
- 0,
1136
- 1
1137
- ]
1138
- ],
1139
- "flatten_loops": [
1140
- true
1141
- ],
1142
- "l2_groupings": [
1143
- 64
1144
- ],
1145
- "range_unroll_factors": [
1146
- 0
1147
- ],
1148
- "range_warp_specializes": [],
1149
- "range_num_stages": [
1150
- 0
1151
- ],
1152
- "range_multi_buffers": [
1153
- null
1154
- ],
1155
- "range_flattens": [
1156
- null
1157
- ],
1158
- "load_eviction_policies": [
1159
- "",
1160
- "",
1161
- ""
1162
- ],
1163
- "num_warps": 2,
1164
- "num_stages": 2,
1165
- "indexing": [
1166
- "pointer",
1167
- "pointer",
1168
- "pointer",
1169
- "pointer"
1170
- ],
1171
- "pid_type": "xyz"
1172
- },
1173
- "intermediate_2048_numtokens_8": {
1174
- "block_sizes": [
1175
- 8,
1176
- 32
1177
- ],
1178
- "loop_orders": [
1179
- [
1180
- 0,
1181
- 1
1182
- ]
1183
- ],
1184
- "flatten_loops": [
1185
- true
1186
- ],
1187
- "l2_groupings": [
1188
- 1
1189
- ],
1190
- "range_unroll_factors": [
1191
- 0
1192
- ],
1193
- "range_warp_specializes": [],
1194
- "range_num_stages": [
1195
- 0
1196
- ],
1197
- "range_multi_buffers": [
1198
- null
1199
- ],
1200
- "range_flattens": [
1201
- null
1202
- ],
1203
- "load_eviction_policies": [
1204
- "",
1205
- "",
1206
- ""
1207
- ],
1208
- "num_warps": 2,
1209
- "num_stages": 1,
1210
- "indexing": [
1211
- "pointer",
1212
- "pointer",
1213
- "pointer",
1214
- "pointer"
1215
- ],
1216
- "pid_type": "flat"
1217
- },
1218
- "intermediate_2880_numtokens_8": {
1219
- "block_sizes": [
1220
- 4,
1221
- 2048
1222
- ],
1223
- "loop_orders": [
1224
- [
1225
- 0,
1226
- 1
1227
- ]
1228
- ],
1229
- "flatten_loops": [
1230
- true
1231
- ],
1232
- "l2_groupings": [
1233
- 4
1234
- ],
1235
- "range_unroll_factors": [
1236
- 0
1237
- ],
1238
- "range_warp_specializes": [],
1239
- "range_num_stages": [
1240
- 0
1241
- ],
1242
- "range_multi_buffers": [
1243
- null
1244
- ],
1245
- "range_flattens": [
1246
- null
1247
- ],
1248
- "load_eviction_policies": [
1249
- "last",
1250
- "last",
1251
- ""
1252
- ],
1253
- "num_warps": 16,
1254
- "num_stages": 5,
1255
- "indexing": [
1256
- "pointer",
1257
- "tensor_descriptor",
1258
- "tensor_descriptor",
1259
- "pointer"
1260
- ],
1261
- "pid_type": "xyz"
1262
- },
1263
- "intermediate_4096_numtokens_8": {
1264
- "block_sizes": [
1265
- 8,
1266
- 32
1267
- ],
1268
- "loop_orders": [
1269
- [
1270
- 0,
1271
- 1
1272
- ]
1273
- ],
1274
- "flatten_loops": [
1275
- true
1276
- ],
1277
- "l2_groupings": [
1278
- 1
1279
- ],
1280
- "range_unroll_factors": [
1281
- 0
1282
- ],
1283
- "range_warp_specializes": [],
1284
- "range_num_stages": [
1285
- 0
1286
- ],
1287
- "range_multi_buffers": [
1288
- null
1289
- ],
1290
- "range_flattens": [
1291
- null
1292
- ],
1293
- "load_eviction_policies": [
1294
- "",
1295
- "",
1296
- ""
1297
- ],
1298
- "num_warps": 2,
1299
- "num_stages": 1,
1300
- "indexing": [
1301
- "pointer",
1302
- "pointer",
1303
- "pointer",
1304
- "pointer"
1305
- ],
1306
- "pid_type": "flat"
1307
- },
1308
- "intermediate_8192_numtokens_8": {
1309
- "block_sizes": [
1310
- 2,
1311
- 1024
1312
- ],
1313
- "loop_orders": [
1314
- [
1315
- 0,
1316
- 1
1317
- ]
1318
- ],
1319
- "flatten_loops": [
1320
- true
1321
- ],
1322
- "l2_groupings": [
1323
- 16
1324
- ],
1325
- "range_unroll_factors": [
1326
- 0
1327
- ],
1328
- "range_warp_specializes": [],
1329
- "range_num_stages": [
1330
- 0
1331
- ],
1332
- "range_multi_buffers": [
1333
- null
1334
- ],
1335
- "range_flattens": [
1336
- null
1337
- ],
1338
- "load_eviction_policies": [
1339
- "first",
1340
- "last",
1341
- ""
1342
- ],
1343
- "num_warps": 1,
1344
- "num_stages": 8,
1345
- "indexing": [
1346
- "tensor_descriptor",
1347
- "pointer",
1348
- "tensor_descriptor",
1349
- "pointer"
1350
- ],
1351
- "pid_type": "flat"
1352
- },
1353
- "intermediate_11008_numtokens_8": {
1354
- "block_sizes": [
1355
- 1,
1356
- 1024
1357
- ],
1358
- "loop_orders": [
1359
- [
1360
- 1,
1361
- 0
1362
- ]
1363
- ],
1364
- "flatten_loops": [
1365
- true
1366
- ],
1367
- "l2_groupings": [
1368
- 4
1369
- ],
1370
- "range_unroll_factors": [
1371
- 0
1372
- ],
1373
- "range_warp_specializes": [],
1374
- "range_num_stages": [
1375
- 0
1376
- ],
1377
- "range_multi_buffers": [
1378
- null
1379
- ],
1380
- "range_flattens": [
1381
- null
1382
- ],
1383
- "load_eviction_policies": [
1384
- "last",
1385
- "",
1386
- "first"
1387
- ],
1388
- "num_warps": 2,
1389
- "num_stages": 5,
1390
- "indexing": [
1391
- "tensor_descriptor",
1392
- "tensor_descriptor",
1393
- "tensor_descriptor",
1394
- "tensor_descriptor"
1395
- ],
1396
- "pid_type": "flat"
1397
- },
1398
- "intermediate_14336_numtokens_8": {
1399
- "block_sizes": [
1400
- 8,
1401
- 128
1402
- ],
1403
- "loop_orders": [
1404
- [
1405
- 0,
1406
- 1
1407
- ]
1408
- ],
1409
- "flatten_loops": [
1410
- true
1411
- ],
1412
- "l2_groupings": [
1413
- 1
1414
- ],
1415
- "range_unroll_factors": [
1416
- 0
1417
- ],
1418
- "range_warp_specializes": [],
1419
- "range_num_stages": [
1420
- 0
1421
- ],
1422
- "range_multi_buffers": [
1423
- null
1424
- ],
1425
- "range_flattens": [
1426
- null
1427
- ],
1428
- "load_eviction_policies": [
1429
- "",
1430
- "",
1431
- ""
1432
- ],
1433
- "num_warps": 4,
1434
- "num_stages": 1,
1435
- "indexing": [
1436
- "pointer",
1437
- "pointer",
1438
- "pointer",
1439
- "pointer"
1440
- ],
1441
- "pid_type": "flat"
1442
- },
1443
- "intermediate_2048_numtokens_16": {
1444
- "block_sizes": [
1445
- 8,
1446
- 512
1447
- ],
1448
- "loop_orders": [
1449
- [
1450
- 0,
1451
- 1
1452
- ]
1453
- ],
1454
- "flatten_loops": [
1455
- true
1456
- ],
1457
- "l2_groupings": [
1458
- 64
1459
- ],
1460
- "range_unroll_factors": [
1461
- 0
1462
- ],
1463
- "range_warp_specializes": [],
1464
- "range_num_stages": [
1465
- 0
1466
- ],
1467
- "range_multi_buffers": [
1468
- null
1469
- ],
1470
- "range_flattens": [
1471
- null
1472
- ],
1473
- "load_eviction_policies": [
1474
- "last",
1475
- "last",
1476
- "first"
1477
- ],
1478
- "num_warps": 16,
1479
- "num_stages": 2,
1480
- "indexing": [
1481
- "pointer",
1482
- "pointer",
1483
- "tensor_descriptor",
1484
- "pointer"
1485
- ],
1486
- "pid_type": "xyz"
1487
- },
1488
- "intermediate_2880_numtokens_16": {
1489
- "block_sizes": [
1490
- 2,
1491
- 256
1492
- ],
1493
- "loop_orders": [
1494
- [
1495
- 0,
1496
- 1
1497
- ]
1498
- ],
1499
- "flatten_loops": [
1500
- true
1501
- ],
1502
- "l2_groupings": [
1503
- 1
1504
- ],
1505
- "range_unroll_factors": [
1506
- 0
1507
- ],
1508
- "range_warp_specializes": [],
1509
- "range_num_stages": [
1510
- 0
1511
- ],
1512
- "range_multi_buffers": [
1513
- null
1514
- ],
1515
- "range_flattens": [
1516
- null
1517
- ],
1518
- "load_eviction_policies": [
1519
- "",
1520
- "",
1521
- ""
1522
- ],
1523
- "num_warps": 4,
1524
- "num_stages": 1,
1525
- "indexing": [
1526
- "pointer",
1527
- "pointer",
1528
- "pointer",
1529
- "pointer"
1530
- ],
1531
- "pid_type": "flat"
1532
- },
1533
- "intermediate_4096_numtokens_16": {
1534
- "block_sizes": [
1535
- 16,
1536
- 256
1537
- ],
1538
- "loop_orders": [
1539
- [
1540
- 0,
1541
- 1
1542
- ]
1543
- ],
1544
- "flatten_loops": [
1545
- true
1546
- ],
1547
- "l2_groupings": [
1548
- 16
1549
- ],
1550
- "range_unroll_factors": [
1551
- 0
1552
- ],
1553
- "range_warp_specializes": [],
1554
- "range_num_stages": [
1555
- 0
1556
- ],
1557
- "range_multi_buffers": [
1558
- null
1559
- ],
1560
- "range_flattens": [
1561
- null
1562
- ],
1563
- "load_eviction_policies": [
1564
- "last",
1565
- "",
1566
- "last"
1567
- ],
1568
- "num_warps": 32,
1569
- "num_stages": 1,
1570
- "indexing": [
1571
- "pointer",
1572
- "pointer",
1573
- "pointer",
1574
- "pointer"
1575
- ],
1576
- "pid_type": "flat"
1577
- },
1578
- "intermediate_8192_numtokens_16": {
1579
- "block_sizes": [
1580
- 16,
1581
- 64
1582
- ],
1583
- "loop_orders": [
1584
- [
1585
- 0,
1586
- 1
1587
- ]
1588
- ],
1589
- "flatten_loops": [
1590
- true
1591
- ],
1592
- "l2_groupings": [
1593
- 1
1594
- ],
1595
- "range_unroll_factors": [
1596
- 0
1597
- ],
1598
- "range_warp_specializes": [],
1599
- "range_num_stages": [
1600
- 0
1601
- ],
1602
- "range_multi_buffers": [
1603
- null
1604
- ],
1605
- "range_flattens": [
1606
- null
1607
- ],
1608
- "load_eviction_policies": [
1609
- "",
1610
- "",
1611
- ""
1612
- ],
1613
- "num_warps": 4,
1614
- "num_stages": 1,
1615
- "indexing": [
1616
- "pointer",
1617
- "pointer",
1618
- "pointer",
1619
- "pointer"
1620
- ],
1621
- "pid_type": "flat"
1622
- },
1623
- "intermediate_11008_numtokens_16": {
1624
- "block_sizes": [
1625
- 1,
1626
- 2048
1627
- ],
1628
- "loop_orders": [
1629
- [
1630
- 0,
1631
- 1
1632
- ]
1633
- ],
1634
- "flatten_loops": [
1635
- true
1636
- ],
1637
- "l2_groupings": [
1638
- 1
1639
- ],
1640
- "range_unroll_factors": [
1641
- 0
1642
- ],
1643
- "range_warp_specializes": [],
1644
- "range_num_stages": [
1645
- 0
1646
- ],
1647
- "range_multi_buffers": [
1648
- null
1649
- ],
1650
- "range_flattens": [
1651
- null
1652
- ],
1653
- "load_eviction_policies": [
1654
- "",
1655
- "",
1656
- ""
1657
- ],
1658
- "num_warps": 8,
1659
- "num_stages": 1,
1660
- "indexing": [
1661
- "pointer",
1662
- "pointer",
1663
- "pointer",
1664
- "pointer"
1665
- ],
1666
- "pid_type": "flat"
1667
- },
1668
- "intermediate_14336_numtokens_16": {
1669
- "block_sizes": [
1670
- 2,
1671
- 256
1672
- ],
1673
- "loop_orders": [
1674
- [
1675
- 1,
1676
- 0
1677
- ]
1678
- ],
1679
- "flatten_loops": [
1680
- true
1681
- ],
1682
- "l2_groupings": [
1683
- 64
1684
- ],
1685
- "range_unroll_factors": [
1686
- 0
1687
- ],
1688
- "range_warp_specializes": [],
1689
- "range_num_stages": [
1690
- 0
1691
- ],
1692
- "range_multi_buffers": [
1693
- null
1694
- ],
1695
- "range_flattens": [
1696
- null
1697
- ],
1698
- "load_eviction_policies": [
1699
- "last",
1700
- "last",
1701
- "last"
1702
- ],
1703
- "num_warps": 16,
1704
- "num_stages": 7,
1705
- "indexing": [
1706
- "tensor_descriptor",
1707
- "pointer",
1708
- "tensor_descriptor",
1709
- "tensor_descriptor"
1710
- ],
1711
- "pid_type": "flat"
1712
- },
1713
- "intermediate_2048_numtokens_24": {
1714
- "block_sizes": [
1715
- 1,
1716
- 1024
1717
- ],
1718
- "loop_orders": [
1719
- [
1720
- 0,
1721
- 1
1722
- ]
1723
- ],
1724
- "flatten_loops": [
1725
- true
1726
- ],
1727
- "l2_groupings": [
1728
- 64
1729
- ],
1730
- "range_unroll_factors": [
1731
- 0
1732
- ],
1733
- "range_warp_specializes": [],
1734
- "range_num_stages": [
1735
- 0
1736
- ],
1737
- "range_multi_buffers": [
1738
- null
1739
- ],
1740
- "range_flattens": [
1741
- null
1742
- ],
1743
- "load_eviction_policies": [
1744
- "",
1745
- "",
1746
- "first"
1747
- ],
1748
- "num_warps": 4,
1749
- "num_stages": 8,
1750
- "indexing": [
1751
- "tensor_descriptor",
1752
- "pointer",
1753
- "pointer",
1754
- "tensor_descriptor"
1755
- ],
1756
- "pid_type": "flat"
1757
- },
1758
- "intermediate_2880_numtokens_24": {
1759
- "block_sizes": [
1760
- 4,
1761
- 1024
1762
- ],
1763
- "loop_orders": [
1764
- [
1765
- 0,
1766
- 1
1767
- ]
1768
- ],
1769
- "flatten_loops": [
1770
- true
1771
- ],
1772
- "l2_groupings": [
1773
- 16
1774
- ],
1775
- "range_unroll_factors": [
1776
- 0
1777
- ],
1778
- "range_warp_specializes": [],
1779
- "range_num_stages": [
1780
- 0
1781
- ],
1782
- "range_multi_buffers": [
1783
- null
1784
- ],
1785
- "range_flattens": [
1786
- null
1787
- ],
1788
- "load_eviction_policies": [
1789
- "first",
1790
- "",
1791
- "first"
1792
- ],
1793
- "num_warps": 16,
1794
- "num_stages": 3,
1795
- "indexing": [
1796
- "tensor_descriptor",
1797
- "pointer",
1798
- "tensor_descriptor",
1799
- "pointer"
1800
- ],
1801
- "pid_type": "flat"
1802
- },
1803
- "intermediate_4096_numtokens_24": {
1804
- "block_sizes": [
1805
- 16,
1806
- 64
1807
- ],
1808
- "loop_orders": [
1809
- [
1810
- 0,
1811
- 1
1812
- ]
1813
- ],
1814
- "flatten_loops": [
1815
- true
1816
- ],
1817
- "l2_groupings": [
1818
- 1
1819
- ],
1820
- "range_unroll_factors": [
1821
- 0
1822
- ],
1823
- "range_warp_specializes": [],
1824
- "range_num_stages": [
1825
- 0
1826
- ],
1827
- "range_multi_buffers": [
1828
- null
1829
- ],
1830
- "range_flattens": [
1831
- null
1832
- ],
1833
- "load_eviction_policies": [
1834
- "",
1835
- "",
1836
- ""
1837
- ],
1838
- "num_warps": 4,
1839
- "num_stages": 1,
1840
- "indexing": [
1841
- "pointer",
1842
- "pointer",
1843
- "pointer",
1844
- "pointer"
1845
- ],
1846
- "pid_type": "flat"
1847
- },
1848
- "intermediate_8192_numtokens_24": {
1849
- "block_sizes": [
1850
- 1,
1851
- 2048
1852
- ],
1853
- "loop_orders": [
1854
- [
1855
- 0,
1856
- 1
1857
- ]
1858
- ],
1859
- "flatten_loops": [
1860
- true
1861
- ],
1862
- "l2_groupings": [
1863
- 32
1864
- ],
1865
- "range_unroll_factors": [
1866
- 0
1867
- ],
1868
- "range_warp_specializes": [],
1869
- "range_num_stages": [
1870
- 0
1871
- ],
1872
- "range_multi_buffers": [
1873
- null
1874
- ],
1875
- "range_flattens": [
1876
- null
1877
- ],
1878
- "load_eviction_policies": [
1879
- "",
1880
- "last",
1881
- "last"
1882
- ],
1883
- "num_warps": 32,
1884
- "num_stages": 5,
1885
- "indexing": [
1886
- "pointer",
1887
- "pointer",
1888
- "pointer",
1889
- "tensor_descriptor"
1890
- ],
1891
- "pid_type": "flat"
1892
- },
1893
- "intermediate_11008_numtokens_24": {
1894
- "block_sizes": [
1895
- 1,
1896
- 1024
1897
- ],
1898
- "loop_orders": [
1899
- [
1900
- 1,
1901
- 0
1902
- ]
1903
- ],
1904
- "flatten_loops": [
1905
- false
1906
- ],
1907
- "l2_groupings": [
1908
- 64
1909
- ],
1910
- "range_unroll_factors": [
1911
- 0
1912
- ],
1913
- "range_warp_specializes": [],
1914
- "range_num_stages": [
1915
- 0
1916
- ],
1917
- "range_multi_buffers": [
1918
- null
1919
- ],
1920
- "range_flattens": [
1921
- null
1922
- ],
1923
- "load_eviction_policies": [
1924
- "last",
1925
- "",
1926
- ""
1927
- ],
1928
- "num_warps": 2,
1929
- "num_stages": 2,
1930
- "indexing": [
1931
- "pointer",
1932
- "pointer",
1933
- "pointer",
1934
- "pointer"
1935
- ],
1936
- "pid_type": "flat"
1937
- },
1938
- "intermediate_14336_numtokens_24": {
1939
- "block_sizes": [
1940
- 8,
1941
- 512
1942
- ],
1943
- "loop_orders": [
1944
- [
1945
- 0,
1946
- 1
1947
- ]
1948
- ],
1949
- "flatten_loops": [
1950
- true
1951
- ],
1952
- "l2_groupings": [
1953
- 4
1954
- ],
1955
- "range_unroll_factors": [
1956
- 0
1957
- ],
1958
- "range_warp_specializes": [],
1959
- "range_num_stages": [
1960
- 0
1961
- ],
1962
- "range_multi_buffers": [
1963
- null
1964
- ],
1965
- "range_flattens": [
1966
- null
1967
- ],
1968
- "load_eviction_policies": [
1969
- "last",
1970
- "first",
1971
- "first"
1972
- ],
1973
- "num_warps": 32,
1974
- "num_stages": 3,
1975
- "indexing": [
1976
- "pointer",
1977
- "tensor_descriptor",
1978
- "pointer",
1979
- "tensor_descriptor"
1980
- ],
1981
- "pid_type": "flat"
1982
- },
1983
- "intermediate_2048_numtokens_32": {
1984
- "block_sizes": [
1985
- 32,
1986
- 16
1987
- ],
1988
- "loop_orders": [
1989
- [
1990
- 0,
1991
- 1
1992
- ]
1993
- ],
1994
- "flatten_loops": [
1995
- true
1996
- ],
1997
- "l2_groupings": [
1998
- 1
1999
- ],
2000
- "range_unroll_factors": [
2001
- 0
2002
- ],
2003
- "range_warp_specializes": [],
2004
- "range_num_stages": [
2005
- 0
2006
- ],
2007
- "range_multi_buffers": [
2008
- null
2009
- ],
2010
- "range_flattens": [
2011
- null
2012
- ],
2013
- "load_eviction_policies": [
2014
- "",
2015
- "",
2016
- ""
2017
- ],
2018
- "num_warps": 4,
2019
- "num_stages": 1,
2020
- "indexing": [
2021
- "pointer",
2022
- "pointer",
2023
- "pointer",
2024
- "pointer"
2025
- ],
2026
- "pid_type": "flat"
2027
- },
2028
- "intermediate_2880_numtokens_32": {
2029
- "block_sizes": [
2030
- 4,
2031
- 256
2032
- ],
2033
- "loop_orders": [
2034
- [
2035
- 0,
2036
- 1
2037
- ]
2038
- ],
2039
- "flatten_loops": [
2040
- true
2041
- ],
2042
- "l2_groupings": [
2043
- 1
2044
- ],
2045
- "range_unroll_factors": [
2046
- 0
2047
- ],
2048
- "range_warp_specializes": [],
2049
- "range_num_stages": [
2050
- 0
2051
- ],
2052
- "range_multi_buffers": [
2053
- null
2054
- ],
2055
- "range_flattens": [
2056
- null
2057
- ],
2058
- "load_eviction_policies": [
2059
- "",
2060
- "",
2061
- ""
2062
- ],
2063
- "num_warps": 4,
2064
- "num_stages": 1,
2065
- "indexing": [
2066
- "pointer",
2067
- "pointer",
2068
- "pointer",
2069
- "pointer"
2070
- ],
2071
- "pid_type": "flat"
2072
- },
2073
- "intermediate_4096_numtokens_32": {
2074
- "block_sizes": [
2075
- 4,
2076
- 4096
2077
- ],
2078
- "loop_orders": [
2079
- [
2080
- 0,
2081
- 1
2082
- ]
2083
- ],
2084
- "flatten_loops": [
2085
- true
2086
- ],
2087
- "l2_groupings": [
2088
- 16
2089
- ],
2090
- "range_unroll_factors": [
2091
- 0
2092
- ],
2093
- "range_warp_specializes": [],
2094
- "range_num_stages": [
2095
- 0
2096
- ],
2097
- "range_multi_buffers": [
2098
- null
2099
- ],
2100
- "range_flattens": [
2101
- null
2102
- ],
2103
- "load_eviction_policies": [
2104
- "first",
2105
- "last",
2106
- ""
2107
- ],
2108
- "num_warps": 32,
2109
- "num_stages": 5,
2110
- "indexing": [
2111
- "pointer",
2112
- "tensor_descriptor",
2113
- "pointer",
2114
- "pointer"
2115
- ],
2116
- "pid_type": "flat"
2117
- },
2118
- "intermediate_8192_numtokens_32": {
2119
- "block_sizes": [
2120
- 4,
2121
- 1024
2122
- ],
2123
- "loop_orders": [
2124
- [
2125
- 0,
2126
- 1
2127
- ]
2128
- ],
2129
- "flatten_loops": [
2130
- true
2131
- ],
2132
- "l2_groupings": [
2133
- 2
2134
- ],
2135
- "range_unroll_factors": [
2136
- 0
2137
- ],
2138
- "range_warp_specializes": [],
2139
- "range_num_stages": [
2140
- 0
2141
- ],
2142
- "range_multi_buffers": [
2143
- null
2144
- ],
2145
- "range_flattens": [
2146
- null
2147
- ],
2148
- "load_eviction_policies": [
2149
- "",
2150
- "last",
2151
- ""
2152
- ],
2153
- "num_warps": 2,
2154
- "num_stages": 3,
2155
- "indexing": [
2156
- "tensor_descriptor",
2157
- "tensor_descriptor",
2158
- "pointer",
2159
- "tensor_descriptor"
2160
- ],
2161
- "pid_type": "flat"
2162
- },
2163
- "intermediate_11008_numtokens_32": {
2164
- "block_sizes": [
2165
- 2,
2166
- 2048
2167
- ],
2168
- "loop_orders": [
2169
- [
2170
- 0,
2171
- 1
2172
- ]
2173
- ],
2174
- "flatten_loops": [
2175
- true
2176
- ],
2177
- "l2_groupings": [
2178
- 1
2179
- ],
2180
- "range_unroll_factors": [
2181
- 0
2182
- ],
2183
- "range_warp_specializes": [],
2184
- "range_num_stages": [
2185
- 0
2186
- ],
2187
- "range_multi_buffers": [
2188
- null
2189
- ],
2190
- "range_flattens": [
2191
- null
2192
- ],
2193
- "load_eviction_policies": [
2194
- "",
2195
- "",
2196
- ""
2197
- ],
2198
- "num_warps": 8,
2199
- "num_stages": 1,
2200
- "indexing": [
2201
- "pointer",
2202
- "pointer",
2203
- "pointer",
2204
- "pointer"
2205
- ],
2206
- "pid_type": "flat"
2207
- },
2208
- "intermediate_14336_numtokens_32": {
2209
- "block_sizes": [
2210
- 1,
2211
- 512
2212
- ],
2213
- "loop_orders": [
2214
- [
2215
- 0,
2216
- 1
2217
- ]
2218
- ],
2219
- "flatten_loops": [
2220
- true
2221
- ],
2222
- "l2_groupings": [
2223
- 64
2224
- ],
2225
- "range_unroll_factors": [
2226
- 0
2227
- ],
2228
- "range_warp_specializes": [],
2229
- "range_num_stages": [
2230
- 0
2231
- ],
2232
- "range_multi_buffers": [
2233
- null
2234
- ],
2235
- "range_flattens": [
2236
- null
2237
- ],
2238
- "load_eviction_policies": [
2239
- "first",
2240
- "",
2241
- ""
2242
- ],
2243
- "num_warps": 4,
2244
- "num_stages": 3,
2245
- "indexing": [
2246
- "pointer",
2247
- "tensor_descriptor",
2248
- "pointer",
2249
- "pointer"
2250
- ],
2251
- "pid_type": "flat"
2252
- },
2253
- "intermediate_2048_numtokens_40": {
2254
- "block_sizes": [
2255
- 32,
2256
- 32
2257
- ],
2258
- "loop_orders": [
2259
- [
2260
- 0,
2261
- 1
2262
- ]
2263
- ],
2264
- "flatten_loops": [
2265
- true
2266
- ],
2267
- "l2_groupings": [
2268
- 1
2269
- ],
2270
- "range_unroll_factors": [
2271
- 0
2272
- ],
2273
- "range_warp_specializes": [],
2274
- "range_num_stages": [
2275
- 0
2276
- ],
2277
- "range_multi_buffers": [
2278
- null
2279
- ],
2280
- "range_flattens": [
2281
- null
2282
- ],
2283
- "load_eviction_policies": [
2284
- "",
2285
- "",
2286
- ""
2287
- ],
2288
- "num_warps": 4,
2289
- "num_stages": 1,
2290
- "indexing": [
2291
- "pointer",
2292
- "pointer",
2293
- "pointer",
2294
- "pointer"
2295
- ],
2296
- "pid_type": "flat"
2297
- },
2298
- "intermediate_2880_numtokens_40": {
2299
- "block_sizes": [
2300
- 1,
2301
- 4096
2302
- ],
2303
- "loop_orders": [
2304
- [
2305
- 1,
2306
- 0
2307
- ]
2308
- ],
2309
- "flatten_loops": [
2310
- false
2311
- ],
2312
- "l2_groupings": [
2313
- 32
2314
- ],
2315
- "range_unroll_factors": [
2316
- 0
2317
- ],
2318
- "range_warp_specializes": [],
2319
- "range_num_stages": [
2320
- 0
2321
- ],
2322
- "range_multi_buffers": [
2323
- null
2324
- ],
2325
- "range_flattens": [
2326
- null
2327
- ],
2328
- "load_eviction_policies": [
2329
- "",
2330
- "last",
2331
- "last"
2332
- ],
2333
- "num_warps": 8,
2334
- "num_stages": 4,
2335
- "indexing": [
2336
- "tensor_descriptor",
2337
- "pointer",
2338
- "pointer",
2339
- "pointer"
2340
- ],
2341
- "pid_type": "flat"
2342
- },
2343
- "intermediate_4096_numtokens_40": {
2344
- "block_sizes": [
2345
- 32,
2346
- 64
2347
- ],
2348
- "loop_orders": [
2349
- [
2350
- 0,
2351
- 1
2352
- ]
2353
- ],
2354
- "flatten_loops": [
2355
- true
2356
- ],
2357
- "l2_groupings": [
2358
- 1
2359
- ],
2360
- "range_unroll_factors": [
2361
- 0
2362
- ],
2363
- "range_warp_specializes": [],
2364
- "range_num_stages": [
2365
- 0
2366
- ],
2367
- "range_multi_buffers": [
2368
- null
2369
- ],
2370
- "range_flattens": [
2371
- null
2372
- ],
2373
- "load_eviction_policies": [
2374
- "",
2375
- "",
2376
- ""
2377
- ],
2378
- "num_warps": 8,
2379
- "num_stages": 1,
2380
- "indexing": [
2381
- "pointer",
2382
- "pointer",
2383
- "pointer",
2384
- "pointer"
2385
- ],
2386
- "pid_type": "flat"
2387
- },
2388
- "intermediate_8192_numtokens_40": {
2389
- "block_sizes": [
2390
- 2,
2391
- 256
2392
- ],
2393
- "loop_orders": [
2394
- [
2395
- 0,
2396
- 1
2397
- ]
2398
- ],
2399
- "flatten_loops": [
2400
- true
2401
- ],
2402
- "l2_groupings": [
2403
- 16
2404
- ],
2405
- "range_unroll_factors": [
2406
- 0
2407
- ],
2408
- "range_warp_specializes": [],
2409
- "range_num_stages": [
2410
- 0
2411
- ],
2412
- "range_multi_buffers": [
2413
- null
2414
- ],
2415
- "range_flattens": [
2416
- null
2417
- ],
2418
- "load_eviction_policies": [
2419
- "",
2420
- "last",
2421
- "last"
2422
- ],
2423
- "num_warps": 2,
2424
- "num_stages": 2,
2425
- "indexing": [
2426
- "pointer",
2427
- "pointer",
2428
- "pointer",
2429
- "tensor_descriptor"
2430
- ],
2431
- "pid_type": "flat"
2432
- },
2433
- "intermediate_11008_numtokens_40": {
2434
- "block_sizes": [
2435
- 16,
2436
- 256
2437
- ],
2438
- "loop_orders": [
2439
- [
2440
- 0,
2441
- 1
2442
- ]
2443
- ],
2444
- "flatten_loops": [
2445
- true
2446
- ],
2447
- "l2_groupings": [
2448
- 1
2449
- ],
2450
- "range_unroll_factors": [
2451
- 0
2452
- ],
2453
- "range_warp_specializes": [],
2454
- "range_num_stages": [
2455
- 0
2456
- ],
2457
- "range_multi_buffers": [
2458
- null
2459
- ],
2460
- "range_flattens": [
2461
- null
2462
- ],
2463
- "load_eviction_policies": [
2464
- "",
2465
- "",
2466
- ""
2467
- ],
2468
- "num_warps": 8,
2469
- "num_stages": 1,
2470
- "indexing": [
2471
- "pointer",
2472
- "pointer",
2473
- "pointer",
2474
- "pointer"
2475
- ],
2476
- "pid_type": "flat"
2477
- },
2478
- "intermediate_14336_numtokens_40": {
2479
- "block_sizes": [
2480
- 4,
2481
- 1024
2482
- ],
2483
- "loop_orders": [
2484
- [
2485
- 0,
2486
- 1
2487
- ]
2488
- ],
2489
- "flatten_loops": [
2490
- true
2491
- ],
2492
- "l2_groupings": [
2493
- 1
2494
- ],
2495
- "range_unroll_factors": [
2496
- 0
2497
- ],
2498
- "range_warp_specializes": [],
2499
- "range_num_stages": [
2500
- 1
2501
- ],
2502
- "range_multi_buffers": [
2503
- null
2504
- ],
2505
- "range_flattens": [
2506
- null
2507
- ],
2508
- "load_eviction_policies": [
2509
- "first",
2510
- "",
2511
- "last"
2512
- ],
2513
- "num_warps": 16,
2514
- "num_stages": 5,
2515
- "indexing": [
2516
- "tensor_descriptor",
2517
- "tensor_descriptor",
2518
- "tensor_descriptor",
2519
- "tensor_descriptor"
2520
- ],
2521
- "pid_type": "persistent_interleaved",
2522
- "num_sm_multiplier": 32,
2523
- "maxnreg": 32
2524
- },
2525
- "intermediate_2048_numtokens_48": {
2526
- "block_sizes": [
2527
- 32,
2528
- 32
2529
- ],
2530
- "loop_orders": [
2531
- [
2532
- 0,
2533
- 1
2534
- ]
2535
- ],
2536
- "flatten_loops": [
2537
- true
2538
- ],
2539
- "l2_groupings": [
2540
- 1
2541
- ],
2542
- "range_unroll_factors": [
2543
- 0
2544
- ],
2545
- "range_warp_specializes": [],
2546
- "range_num_stages": [
2547
- 0
2548
- ],
2549
- "range_multi_buffers": [
2550
- null
2551
- ],
2552
- "range_flattens": [
2553
- null
2554
- ],
2555
- "load_eviction_policies": [
2556
- "",
2557
- "",
2558
- ""
2559
- ],
2560
- "num_warps": 4,
2561
- "num_stages": 1,
2562
- "indexing": [
2563
- "pointer",
2564
- "pointer",
2565
- "pointer",
2566
- "pointer"
2567
- ],
2568
- "pid_type": "flat"
2569
- },
2570
- "intermediate_2880_numtokens_48": {
2571
- "block_sizes": [
2572
- 16,
2573
- 64
2574
- ],
2575
- "loop_orders": [
2576
- [
2577
- 0,
2578
- 1
2579
- ]
2580
- ],
2581
- "flatten_loops": [
2582
- true
2583
- ],
2584
- "l2_groupings": [
2585
- 1
2586
- ],
2587
- "range_unroll_factors": [
2588
- 0
2589
- ],
2590
- "range_warp_specializes": [],
2591
- "range_num_stages": [
2592
- 0
2593
- ],
2594
- "range_multi_buffers": [
2595
- null
2596
- ],
2597
- "range_flattens": [
2598
- null
2599
- ],
2600
- "load_eviction_policies": [
2601
- "",
2602
- "",
2603
- ""
2604
- ],
2605
- "num_warps": 4,
2606
- "num_stages": 1,
2607
- "indexing": [
2608
- "pointer",
2609
- "pointer",
2610
- "pointer",
2611
- "pointer"
2612
- ],
2613
- "pid_type": "flat"
2614
- },
2615
- "intermediate_4096_numtokens_48": {
2616
- "block_sizes": [
2617
- 32,
2618
- 64
2619
- ],
2620
- "loop_orders": [
2621
- [
2622
- 0,
2623
- 1
2624
- ]
2625
- ],
2626
- "flatten_loops": [
2627
- true
2628
- ],
2629
- "l2_groupings": [
2630
- 1
2631
- ],
2632
- "range_unroll_factors": [
2633
- 0
2634
- ],
2635
- "range_warp_specializes": [],
2636
- "range_num_stages": [
2637
- 0
2638
- ],
2639
- "range_multi_buffers": [
2640
- null
2641
- ],
2642
- "range_flattens": [
2643
- null
2644
- ],
2645
- "load_eviction_policies": [
2646
- "",
2647
- "",
2648
- ""
2649
- ],
2650
- "num_warps": 8,
2651
- "num_stages": 1,
2652
- "indexing": [
2653
- "pointer",
2654
- "pointer",
2655
- "pointer",
2656
- "pointer"
2657
- ],
2658
- "pid_type": "flat"
2659
- },
2660
- "intermediate_8192_numtokens_48": {
2661
- "block_sizes": [
2662
- 1,
2663
- 4096
2664
- ],
2665
- "loop_orders": [
2666
- [
2667
- 0,
2668
- 1
2669
- ]
2670
- ],
2671
- "flatten_loops": [
2672
- true
2673
- ],
2674
- "l2_groupings": [
2675
- 2
2676
- ],
2677
- "range_unroll_factors": [
2678
- 0
2679
- ],
2680
- "range_warp_specializes": [],
2681
- "range_num_stages": [
2682
- 0
2683
- ],
2684
- "range_multi_buffers": [
2685
- null
2686
- ],
2687
- "range_flattens": [
2688
- null
2689
- ],
2690
- "load_eviction_policies": [
2691
- "first",
2692
- "",
2693
- "last"
2694
- ],
2695
- "num_warps": 4,
2696
- "num_stages": 2,
2697
- "indexing": [
2698
- "pointer",
2699
- "tensor_descriptor",
2700
- "tensor_descriptor",
2701
- "pointer"
2702
- ],
2703
- "pid_type": "flat"
2704
- },
2705
- "intermediate_11008_numtokens_48": {
2706
- "block_sizes": [
2707
- 16,
2708
- 256
2709
- ],
2710
- "loop_orders": [
2711
- [
2712
- 0,
2713
- 1
2714
- ]
2715
- ],
2716
- "flatten_loops": [
2717
- true
2718
- ],
2719
- "l2_groupings": [
2720
- 1
2721
- ],
2722
- "range_unroll_factors": [
2723
- 0
2724
- ],
2725
- "range_warp_specializes": [],
2726
- "range_num_stages": [
2727
- 0
2728
- ],
2729
- "range_multi_buffers": [
2730
- null
2731
- ],
2732
- "range_flattens": [
2733
- null
2734
- ],
2735
- "load_eviction_policies": [
2736
- "",
2737
- "",
2738
- ""
2739
- ],
2740
- "num_warps": 8,
2741
- "num_stages": 1,
2742
- "indexing": [
2743
- "pointer",
2744
- "pointer",
2745
- "pointer",
2746
- "pointer"
2747
- ],
2748
- "pid_type": "flat"
2749
- },
2750
- "intermediate_14336_numtokens_48": {
2751
- "block_sizes": [
2752
- 32,
2753
- 256
2754
- ],
2755
- "loop_orders": [
2756
- [
2757
- 0,
2758
- 1
2759
- ]
2760
- ],
2761
- "flatten_loops": [
2762
- true
2763
- ],
2764
- "l2_groupings": [
2765
- 1
2766
- ],
2767
- "range_unroll_factors": [
2768
- 0
2769
- ],
2770
- "range_warp_specializes": [],
2771
- "range_num_stages": [
2772
- 0
2773
- ],
2774
- "range_multi_buffers": [
2775
- null
2776
- ],
2777
- "range_flattens": [
2778
- null
2779
- ],
2780
- "load_eviction_policies": [
2781
- "",
2782
- "",
2783
- ""
2784
- ],
2785
- "num_warps": 8,
2786
- "num_stages": 1,
2787
- "indexing": [
2788
- "pointer",
2789
- "pointer",
2790
- "pointer",
2791
- "pointer"
2792
- ],
2793
- "pid_type": "flat"
2794
- },
2795
- "intermediate_2048_numtokens_56": {
2796
- "block_sizes": [
2797
- 32,
2798
- 32
2799
- ],
2800
- "loop_orders": [
2801
- [
2802
- 0,
2803
- 1
2804
- ]
2805
- ],
2806
- "flatten_loops": [
2807
- true
2808
- ],
2809
- "l2_groupings": [
2810
- 1
2811
- ],
2812
- "range_unroll_factors": [
2813
- 0
2814
- ],
2815
- "range_warp_specializes": [],
2816
- "range_num_stages": [
2817
- 0
2818
- ],
2819
- "range_multi_buffers": [
2820
- null
2821
- ],
2822
- "range_flattens": [
2823
- null
2824
- ],
2825
- "load_eviction_policies": [
2826
- "",
2827
- "",
2828
- ""
2829
- ],
2830
- "num_warps": 4,
2831
- "num_stages": 1,
2832
- "indexing": [
2833
- "pointer",
2834
- "pointer",
2835
- "pointer",
2836
- "pointer"
2837
- ],
2838
- "pid_type": "flat"
2839
- },
2840
- "intermediate_2880_numtokens_56": {
2841
- "block_sizes": [
2842
- 1,
2843
- 2048
2844
- ],
2845
- "loop_orders": [
2846
- [
2847
- 0,
2848
- 1
2849
- ]
2850
- ],
2851
- "flatten_loops": [
2852
- true
2853
- ],
2854
- "l2_groupings": [
2855
- 1
2856
- ],
2857
- "range_unroll_factors": [
2858
- 0
2859
- ],
2860
- "range_warp_specializes": [],
2861
- "range_num_stages": [
2862
- 0
2863
- ],
2864
- "range_multi_buffers": [
2865
- null
2866
- ],
2867
- "range_flattens": [
2868
- null
2869
- ],
2870
- "load_eviction_policies": [
2871
- "",
2872
- "",
2873
- ""
2874
- ],
2875
- "num_warps": 8,
2876
- "num_stages": 1,
2877
- "indexing": [
2878
- "pointer",
2879
- "pointer",
2880
- "pointer",
2881
- "pointer"
2882
- ],
2883
- "pid_type": "flat"
2884
- },
2885
- "intermediate_4096_numtokens_56": {
2886
- "block_sizes": [
2887
- 32,
2888
- 64
2889
- ],
2890
- "loop_orders": [
2891
- [
2892
- 0,
2893
- 1
2894
- ]
2895
- ],
2896
- "flatten_loops": [
2897
- true
2898
- ],
2899
- "l2_groupings": [
2900
- 1
2901
- ],
2902
- "range_unroll_factors": [
2903
- 0
2904
- ],
2905
- "range_warp_specializes": [],
2906
- "range_num_stages": [
2907
- 0
2908
- ],
2909
- "range_multi_buffers": [
2910
- null
2911
- ],
2912
- "range_flattens": [
2913
- null
2914
- ],
2915
- "load_eviction_policies": [
2916
- "",
2917
- "",
2918
- ""
2919
- ],
2920
- "num_warps": 8,
2921
- "num_stages": 1,
2922
- "indexing": [
2923
- "pointer",
2924
- "pointer",
2925
- "pointer",
2926
- "pointer"
2927
- ],
2928
- "pid_type": "flat"
2929
- },
2930
- "intermediate_8192_numtokens_56": {
2931
- "block_sizes": [
2932
- 32,
2933
- 128
2934
- ],
2935
- "loop_orders": [
2936
- [
2937
- 0,
2938
- 1
2939
- ]
2940
- ],
2941
- "flatten_loops": [
2942
- true
2943
- ],
2944
- "l2_groupings": [
2945
- 1
2946
- ],
2947
- "range_unroll_factors": [
2948
- 0
2949
- ],
2950
- "range_warp_specializes": [],
2951
- "range_num_stages": [
2952
- 0
2953
- ],
2954
- "range_multi_buffers": [
2955
- null
2956
- ],
2957
- "range_flattens": [
2958
- null
2959
- ],
2960
- "load_eviction_policies": [
2961
- "",
2962
- "",
2963
- ""
2964
- ],
2965
- "num_warps": 8,
2966
- "num_stages": 1,
2967
- "indexing": [
2968
- "pointer",
2969
- "pointer",
2970
- "pointer",
2971
- "pointer"
2972
- ],
2973
- "pid_type": "flat"
2974
- },
2975
- "intermediate_11008_numtokens_56": {
2976
- "block_sizes": [
2977
- 1,
2978
- 8192
2979
- ],
2980
- "loop_orders": [
2981
- [
2982
- 0,
2983
- 1
2984
- ]
2985
- ],
2986
- "flatten_loops": [
2987
- true
2988
- ],
2989
- "l2_groupings": [
2990
- 1
2991
- ],
2992
- "range_unroll_factors": [
2993
- 0
2994
- ],
2995
- "range_warp_specializes": [],
2996
- "range_num_stages": [
2997
- 0
2998
- ],
2999
- "range_multi_buffers": [
3000
- null
3001
- ],
3002
- "range_flattens": [
3003
- null
3004
- ],
3005
- "load_eviction_policies": [
3006
- "",
3007
- "",
3008
- ""
3009
- ],
3010
- "num_warps": 8,
3011
- "num_stages": 1,
3012
- "indexing": [
3013
- "pointer",
3014
- "pointer",
3015
- "pointer",
3016
- "pointer"
3017
- ],
3018
- "pid_type": "flat"
3019
- },
3020
- "intermediate_14336_numtokens_56": {
3021
- "block_sizes": [
3022
- 2,
3023
- 4096
3024
- ],
3025
- "loop_orders": [
3026
- [
3027
- 0,
3028
- 1
3029
- ]
3030
- ],
3031
- "flatten_loops": [
3032
- false
3033
- ],
3034
- "l2_groupings": [
3035
- 32
3036
- ],
3037
- "range_unroll_factors": [
3038
- 0
3039
- ],
3040
- "range_warp_specializes": [],
3041
- "range_num_stages": [
3042
- 0
3043
- ],
3044
- "range_multi_buffers": [
3045
- null
3046
- ],
3047
- "range_flattens": [
3048
- null
3049
- ],
3050
- "load_eviction_policies": [
3051
- "first",
3052
- "first",
3053
- ""
3054
- ],
3055
- "num_warps": 2,
3056
- "num_stages": 4,
3057
- "indexing": [
3058
- "tensor_descriptor",
3059
- "pointer",
3060
- "tensor_descriptor",
3061
- "tensor_descriptor"
3062
- ],
3063
- "pid_type": "flat"
3064
- },
3065
- "intermediate_2048_numtokens_64": {
3066
- "block_sizes": [
3067
- 64,
3068
- 32
3069
- ],
3070
- "loop_orders": [
3071
- [
3072
- 0,
3073
- 1
3074
- ]
3075
- ],
3076
- "flatten_loops": [
3077
- true
3078
- ],
3079
- "l2_groupings": [
3080
- 1
3081
- ],
3082
- "range_unroll_factors": [
3083
- 0
3084
- ],
3085
- "range_warp_specializes": [],
3086
- "range_num_stages": [
3087
- 0
3088
- ],
3089
- "range_multi_buffers": [
3090
- null
3091
- ],
3092
- "range_flattens": [
3093
- null
3094
- ],
3095
- "load_eviction_policies": [
3096
- "",
3097
- "",
3098
- ""
3099
- ],
3100
- "num_warps": 8,
3101
- "num_stages": 1,
3102
- "indexing": [
3103
- "pointer",
3104
- "pointer",
3105
- "pointer",
3106
- "pointer"
3107
- ],
3108
- "pid_type": "flat"
3109
- },
3110
- "intermediate_2880_numtokens_64": {
3111
- "block_sizes": [
3112
- 1,
3113
- 2048
3114
- ],
3115
- "loop_orders": [
3116
- [
3117
- 0,
3118
- 1
3119
- ]
3120
- ],
3121
- "flatten_loops": [
3122
- true
3123
- ],
3124
- "l2_groupings": [
3125
- 1
3126
- ],
3127
- "range_unroll_factors": [
3128
- 0
3129
- ],
3130
- "range_warp_specializes": [],
3131
- "range_num_stages": [
3132
- 0
3133
- ],
3134
- "range_multi_buffers": [
3135
- null
3136
- ],
3137
- "range_flattens": [
3138
- null
3139
- ],
3140
- "load_eviction_policies": [
3141
- "",
3142
- "",
3143
- ""
3144
- ],
3145
- "num_warps": 8,
3146
- "num_stages": 1,
3147
- "indexing": [
3148
- "pointer",
3149
- "pointer",
3150
- "pointer",
3151
- "pointer"
3152
- ],
3153
- "pid_type": "flat"
3154
- },
3155
- "intermediate_4096_numtokens_64": {
3156
- "block_sizes": [
3157
- 64,
3158
- 32
3159
- ],
3160
- "loop_orders": [
3161
- [
3162
- 0,
3163
- 1
3164
- ]
3165
- ],
3166
- "flatten_loops": [
3167
- true
3168
- ],
3169
- "l2_groupings": [
3170
- 1
3171
- ],
3172
- "range_unroll_factors": [
3173
- 0
3174
- ],
3175
- "range_warp_specializes": [],
3176
- "range_num_stages": [
3177
- 0
3178
- ],
3179
- "range_multi_buffers": [
3180
- null
3181
- ],
3182
- "range_flattens": [
3183
- null
3184
- ],
3185
- "load_eviction_policies": [
3186
- "",
3187
- "",
3188
- ""
3189
- ],
3190
- "num_warps": 8,
3191
- "num_stages": 1,
3192
- "indexing": [
3193
- "pointer",
3194
- "pointer",
3195
- "pointer",
3196
- "pointer"
3197
- ],
3198
- "pid_type": "flat"
3199
- },
3200
- "intermediate_8192_numtokens_64": {
3201
- "block_sizes": [
3202
- 64,
3203
- 64
3204
- ],
3205
- "loop_orders": [
3206
- [
3207
- 0,
3208
- 1
3209
- ]
3210
- ],
3211
- "flatten_loops": [
3212
- true
3213
- ],
3214
- "l2_groupings": [
3215
- 1
3216
- ],
3217
- "range_unroll_factors": [
3218
- 0
3219
- ],
3220
- "range_warp_specializes": [],
3221
- "range_num_stages": [
3222
- 0
3223
- ],
3224
- "range_multi_buffers": [
3225
- null
3226
- ],
3227
- "range_flattens": [
3228
- null
3229
- ],
3230
- "load_eviction_policies": [
3231
- "",
3232
- "",
3233
- ""
3234
- ],
3235
- "num_warps": 8,
3236
- "num_stages": 1,
3237
- "indexing": [
3238
- "pointer",
3239
- "pointer",
3240
- "pointer",
3241
- "pointer"
3242
- ],
3243
- "pid_type": "flat"
3244
- },
3245
- "intermediate_11008_numtokens_64": {
3246
- "block_sizes": [
3247
- 1,
3248
- 8192
3249
- ],
3250
- "loop_orders": [
3251
- [
3252
- 0,
3253
- 1
3254
- ]
3255
- ],
3256
- "flatten_loops": [
3257
- true
3258
- ],
3259
- "l2_groupings": [
3260
- 1
3261
- ],
3262
- "range_unroll_factors": [
3263
- 0
3264
- ],
3265
- "range_warp_specializes": [],
3266
- "range_num_stages": [
3267
- 0
3268
- ],
3269
- "range_multi_buffers": [
3270
- null
3271
- ],
3272
- "range_flattens": [
3273
- null
3274
- ],
3275
- "load_eviction_policies": [
3276
- "",
3277
- "",
3278
- ""
3279
- ],
3280
- "num_warps": 8,
3281
- "num_stages": 1,
3282
- "indexing": [
3283
- "pointer",
3284
- "pointer",
3285
- "pointer",
3286
- "pointer"
3287
- ],
3288
- "pid_type": "flat"
3289
- },
3290
- "intermediate_14336_numtokens_64": {
3291
- "block_sizes": [
3292
- 16,
3293
- 512
3294
- ],
3295
- "loop_orders": [
3296
- [
3297
- 0,
3298
- 1
3299
- ]
3300
- ],
3301
- "flatten_loops": [
3302
- true
3303
- ],
3304
- "l2_groupings": [
3305
- 16
3306
- ],
3307
- "range_unroll_factors": [
3308
- 0
3309
- ],
3310
- "range_warp_specializes": [],
3311
- "range_num_stages": [
3312
- 0
3313
- ],
3314
- "range_multi_buffers": [
3315
- null
3316
- ],
3317
- "range_flattens": [
3318
- null
3319
- ],
3320
- "load_eviction_policies": [
3321
- "",
3322
- "first",
3323
- ""
3324
- ],
3325
- "num_warps": 4,
3326
- "num_stages": 1,
3327
- "indexing": [
3328
- "pointer",
3329
- "tensor_descriptor",
3330
- "tensor_descriptor",
3331
- "tensor_descriptor"
3332
- ],
3333
- "pid_type": "flat"
3334
- },
3335
- "intermediate_2048_numtokens_72": {
3336
- "block_sizes": [
3337
- 64,
3338
- 32
3339
- ],
3340
- "loop_orders": [
3341
- [
3342
- 0,
3343
- 1
3344
- ]
3345
- ],
3346
- "flatten_loops": [
3347
- true
3348
- ],
3349
- "l2_groupings": [
3350
- 1
3351
- ],
3352
- "range_unroll_factors": [
3353
- 0
3354
- ],
3355
- "range_warp_specializes": [],
3356
- "range_num_stages": [
3357
- 0
3358
- ],
3359
- "range_multi_buffers": [
3360
- null
3361
- ],
3362
- "range_flattens": [
3363
- null
3364
- ],
3365
- "load_eviction_policies": [
3366
- "",
3367
- "",
3368
- ""
3369
- ],
3370
- "num_warps": 8,
3371
- "num_stages": 1,
3372
- "indexing": [
3373
- "pointer",
3374
- "pointer",
3375
- "pointer",
3376
- "pointer"
3377
- ],
3378
- "pid_type": "flat"
3379
- },
3380
- "intermediate_2880_numtokens_72": {
3381
- "block_sizes": [
3382
- 32,
3383
- 64
3384
- ],
3385
- "loop_orders": [
3386
- [
3387
- 0,
3388
- 1
3389
- ]
3390
- ],
3391
- "flatten_loops": [
3392
- true
3393
- ],
3394
- "l2_groupings": [
3395
- 1
3396
- ],
3397
- "range_unroll_factors": [
3398
- 0
3399
- ],
3400
- "range_warp_specializes": [],
3401
- "range_num_stages": [
3402
- 0
3403
- ],
3404
- "range_multi_buffers": [
3405
- null
3406
- ],
3407
- "range_flattens": [
3408
- null
3409
- ],
3410
- "load_eviction_policies": [
3411
- "",
3412
- "",
3413
- ""
3414
- ],
3415
- "num_warps": 8,
3416
- "num_stages": 1,
3417
- "indexing": [
3418
- "pointer",
3419
- "pointer",
3420
- "pointer",
3421
- "pointer"
3422
- ],
3423
- "pid_type": "flat"
3424
- },
3425
- "intermediate_4096_numtokens_72": {
3426
- "block_sizes": [
3427
- 1,
3428
- 1024
3429
- ],
3430
- "loop_orders": [
3431
- [
3432
- 0,
3433
- 1
3434
- ]
3435
- ],
3436
- "flatten_loops": [
3437
- true
3438
- ],
3439
- "l2_groupings": [
3440
- 4
3441
- ],
3442
- "range_unroll_factors": [
3443
- 0
3444
- ],
3445
- "range_warp_specializes": [],
3446
- "range_num_stages": [
3447
- 0
3448
- ],
3449
- "range_multi_buffers": [
3450
- null
3451
- ],
3452
- "range_flattens": [
3453
- null
3454
- ],
3455
- "load_eviction_policies": [
3456
- "",
3457
- "",
3458
- "first"
3459
- ],
3460
- "num_warps": 16,
3461
- "num_stages": 2,
3462
- "indexing": [
3463
- "tensor_descriptor",
3464
- "pointer",
3465
- "pointer",
3466
- "tensor_descriptor"
3467
- ],
3468
- "pid_type": "flat"
3469
- },
3470
- "intermediate_8192_numtokens_72": {
3471
- "block_sizes": [
3472
- 64,
3473
- 128
3474
- ],
3475
- "loop_orders": [
3476
- [
3477
- 0,
3478
- 1
3479
- ]
3480
- ],
3481
- "flatten_loops": [
3482
- true
3483
- ],
3484
- "l2_groupings": [
3485
- 1
3486
- ],
3487
- "range_unroll_factors": [
3488
- 0
3489
- ],
3490
- "range_warp_specializes": [],
3491
- "range_num_stages": [
3492
- 0
3493
- ],
3494
- "range_multi_buffers": [
3495
- null
3496
- ],
3497
- "range_flattens": [
3498
- null
3499
- ],
3500
- "load_eviction_policies": [
3501
- "",
3502
- "",
3503
- ""
3504
- ],
3505
- "num_warps": 8,
3506
- "num_stages": 1,
3507
- "indexing": [
3508
- "pointer",
3509
- "pointer",
3510
- "pointer",
3511
- "pointer"
3512
- ],
3513
- "pid_type": "flat"
3514
- },
3515
- "intermediate_11008_numtokens_72": {
3516
- "block_sizes": [
3517
- 4,
3518
- 256
3519
- ],
3520
- "loop_orders": [
3521
- [
3522
- 0,
3523
- 1
3524
- ]
3525
- ],
3526
- "flatten_loops": [
3527
- true
3528
- ],
3529
- "l2_groupings": [
3530
- 2
3531
- ],
3532
- "range_unroll_factors": [
3533
- 0
3534
- ],
3535
- "range_warp_specializes": [],
3536
- "range_num_stages": [
3537
- 0
3538
- ],
3539
- "range_multi_buffers": [
3540
- null
3541
- ],
3542
- "range_flattens": [
3543
- null
3544
- ],
3545
- "load_eviction_policies": [
3546
- "first",
3547
- "last",
3548
- "last"
3549
- ],
3550
- "num_warps": 32,
3551
- "num_stages": 5,
3552
- "indexing": [
3553
- "pointer",
3554
- "pointer",
3555
- "tensor_descriptor",
3556
- "tensor_descriptor"
3557
- ],
3558
- "pid_type": "flat"
3559
- },
3560
- "intermediate_14336_numtokens_72": {
3561
- "block_sizes": [
3562
- 32,
3563
- 32
3564
- ],
3565
- "loop_orders": [
3566
- [
3567
- 0,
3568
- 1
3569
- ]
3570
- ],
3571
- "flatten_loops": [
3572
- true
3573
- ],
3574
- "l2_groupings": [
3575
- 2
3576
- ],
3577
- "range_unroll_factors": [
3578
- 0
3579
- ],
3580
- "range_warp_specializes": [],
3581
- "range_num_stages": [
3582
- 0
3583
- ],
3584
- "range_multi_buffers": [
3585
- null
3586
- ],
3587
- "range_flattens": [
3588
- null
3589
- ],
3590
- "load_eviction_policies": [
3591
- "last",
3592
- "",
3593
- "last"
3594
- ],
3595
- "num_warps": 8,
3596
- "num_stages": 3,
3597
- "indexing": [
3598
- "pointer",
3599
- "pointer",
3600
- "tensor_descriptor",
3601
- "pointer"
3602
- ],
3603
- "pid_type": "flat"
3604
- },
3605
- "intermediate_2048_numtokens_80": {
3606
- "block_sizes": [
3607
- 64,
3608
- 32
3609
- ],
3610
- "loop_orders": [
3611
- [
3612
- 0,
3613
- 1
3614
- ]
3615
- ],
3616
- "flatten_loops": [
3617
- true
3618
- ],
3619
- "l2_groupings": [
3620
- 1
3621
- ],
3622
- "range_unroll_factors": [
3623
- 0
3624
- ],
3625
- "range_warp_specializes": [],
3626
- "range_num_stages": [
3627
- 0
3628
- ],
3629
- "range_multi_buffers": [
3630
- null
3631
- ],
3632
- "range_flattens": [
3633
- null
3634
- ],
3635
- "load_eviction_policies": [
3636
- "",
3637
- "",
3638
- ""
3639
- ],
3640
- "num_warps": 8,
3641
- "num_stages": 1,
3642
- "indexing": [
3643
- "pointer",
3644
- "pointer",
3645
- "pointer",
3646
- "pointer"
3647
- ],
3648
- "pid_type": "flat"
3649
- },
3650
- "intermediate_2880_numtokens_80": {
3651
- "block_sizes": [
3652
- 32,
3653
- 64
3654
- ],
3655
- "loop_orders": [
3656
- [
3657
- 0,
3658
- 1
3659
- ]
3660
- ],
3661
- "flatten_loops": [
3662
- true
3663
- ],
3664
- "l2_groupings": [
3665
- 1
3666
- ],
3667
- "range_unroll_factors": [
3668
- 0
3669
- ],
3670
- "range_warp_specializes": [],
3671
- "range_num_stages": [
3672
- 0
3673
- ],
3674
- "range_multi_buffers": [
3675
- null
3676
- ],
3677
- "range_flattens": [
3678
- null
3679
- ],
3680
- "load_eviction_policies": [
3681
- "",
3682
- "",
3683
- ""
3684
- ],
3685
- "num_warps": 8,
3686
- "num_stages": 1,
3687
- "indexing": [
3688
- "pointer",
3689
- "pointer",
3690
- "pointer",
3691
- "pointer"
3692
- ],
3693
- "pid_type": "flat"
3694
- },
3695
- "intermediate_4096_numtokens_80": {
3696
- "block_sizes": [
3697
- 64,
3698
- 64
3699
- ],
3700
- "loop_orders": [
3701
- [
3702
- 0,
3703
- 1
3704
- ]
3705
- ],
3706
- "flatten_loops": [
3707
- true
3708
- ],
3709
- "l2_groupings": [
3710
- 1
3711
- ],
3712
- "range_unroll_factors": [
3713
- 0
3714
- ],
3715
- "range_warp_specializes": [],
3716
- "range_num_stages": [
3717
- 0
3718
- ],
3719
- "range_multi_buffers": [
3720
- null
3721
- ],
3722
- "range_flattens": [
3723
- null
3724
- ],
3725
- "load_eviction_policies": [
3726
- "",
3727
- "",
3728
- ""
3729
- ],
3730
- "num_warps": 8,
3731
- "num_stages": 1,
3732
- "indexing": [
3733
- "pointer",
3734
- "pointer",
3735
- "pointer",
3736
- "pointer"
3737
- ],
3738
- "pid_type": "flat"
3739
- },
3740
- "intermediate_8192_numtokens_80": {
3741
- "block_sizes": [
3742
- 4,
3743
- 128
3744
- ],
3745
- "loop_orders": [
3746
- [
3747
- 0,
3748
- 1
3749
- ]
3750
- ],
3751
- "flatten_loops": [
3752
- true
3753
- ],
3754
- "l2_groupings": [
3755
- 4
3756
- ],
3757
- "range_unroll_factors": [
3758
- 0
3759
- ],
3760
- "range_warp_specializes": [],
3761
- "range_num_stages": [
3762
- 0
3763
- ],
3764
- "range_multi_buffers": [
3765
- null
3766
- ],
3767
- "range_flattens": [
3768
- null
3769
- ],
3770
- "load_eviction_policies": [
3771
- "last",
3772
- "last",
3773
- "last"
3774
- ],
3775
- "num_warps": 16,
3776
- "num_stages": 2,
3777
- "indexing": [
3778
- "pointer",
3779
- "pointer",
3780
- "pointer",
3781
- "pointer"
3782
- ],
3783
- "pid_type": "flat"
3784
- },
3785
- "intermediate_11008_numtokens_80": {
3786
- "block_sizes": [
3787
- 1,
3788
- 4096
3789
- ],
3790
- "loop_orders": [
3791
- [
3792
- 0,
3793
- 1
3794
- ]
3795
- ],
3796
- "flatten_loops": [
3797
- true
3798
- ],
3799
- "l2_groupings": [
3800
- 32
3801
- ],
3802
- "range_unroll_factors": [
3803
- 0
3804
- ],
3805
- "range_warp_specializes": [],
3806
- "range_num_stages": [
3807
- 0
3808
- ],
3809
- "range_multi_buffers": [
3810
- null
3811
- ],
3812
- "range_flattens": [
3813
- null
3814
- ],
3815
- "load_eviction_policies": [
3816
- "last",
3817
- "last",
3818
- ""
3819
- ],
3820
- "num_warps": 16,
3821
- "num_stages": 2,
3822
- "indexing": [
3823
- "pointer",
3824
- "tensor_descriptor",
3825
- "tensor_descriptor",
3826
- "tensor_descriptor"
3827
- ],
3828
- "pid_type": "flat"
3829
- },
3830
- "intermediate_14336_numtokens_80": {
3831
- "block_sizes": [
3832
- 2,
3833
- 2048
3834
- ],
3835
- "loop_orders": [
3836
- [
3837
- 0,
3838
- 1
3839
- ]
3840
- ],
3841
- "flatten_loops": [
3842
- true
3843
- ],
3844
- "l2_groupings": [
3845
- 4
3846
- ],
3847
- "range_unroll_factors": [
3848
- 0
3849
- ],
3850
- "range_warp_specializes": [],
3851
- "range_num_stages": [
3852
- 0
3853
- ],
3854
- "range_multi_buffers": [
3855
- null
3856
- ],
3857
- "range_flattens": [
3858
- null
3859
- ],
3860
- "load_eviction_policies": [
3861
- "last",
3862
- "last",
3863
- ""
3864
- ],
3865
- "num_warps": 4,
3866
- "num_stages": 6,
3867
- "indexing": [
3868
- "tensor_descriptor",
3869
- "pointer",
3870
- "pointer",
3871
- "tensor_descriptor"
3872
- ],
3873
- "pid_type": "flat"
3874
- },
3875
- "intermediate_2048_numtokens_88": {
3876
- "block_sizes": [
3877
- 64,
3878
- 32
3879
- ],
3880
- "loop_orders": [
3881
- [
3882
- 0,
3883
- 1
3884
- ]
3885
- ],
3886
- "flatten_loops": [
3887
- true
3888
- ],
3889
- "l2_groupings": [
3890
- 1
3891
- ],
3892
- "range_unroll_factors": [
3893
- 0
3894
- ],
3895
- "range_warp_specializes": [],
3896
- "range_num_stages": [
3897
- 0
3898
- ],
3899
- "range_multi_buffers": [
3900
- null
3901
- ],
3902
- "range_flattens": [
3903
- null
3904
- ],
3905
- "load_eviction_policies": [
3906
- "",
3907
- "",
3908
- ""
3909
- ],
3910
- "num_warps": 8,
3911
- "num_stages": 1,
3912
- "indexing": [
3913
- "pointer",
3914
- "pointer",
3915
- "pointer",
3916
- "pointer"
3917
- ],
3918
- "pid_type": "flat"
3919
- },
3920
- "intermediate_2880_numtokens_88": {
3921
- "block_sizes": [
3922
- 8,
3923
- 256
3924
- ],
3925
- "loop_orders": [
3926
- [
3927
- 0,
3928
- 1
3929
- ]
3930
- ],
3931
- "flatten_loops": [
3932
- true
3933
- ],
3934
- "l2_groupings": [
3935
- 1
3936
- ],
3937
- "range_unroll_factors": [
3938
- 0
3939
- ],
3940
- "range_warp_specializes": [],
3941
- "range_num_stages": [
3942
- 0
3943
- ],
3944
- "range_multi_buffers": [
3945
- null
3946
- ],
3947
- "range_flattens": [
3948
- null
3949
- ],
3950
- "load_eviction_policies": [
3951
- "",
3952
- "",
3953
- ""
3954
- ],
3955
- "num_warps": 8,
3956
- "num_stages": 1,
3957
- "indexing": [
3958
- "pointer",
3959
- "pointer",
3960
- "pointer",
3961
- "pointer"
3962
- ],
3963
- "pid_type": "flat"
3964
- },
3965
- "intermediate_4096_numtokens_88": {
3966
- "block_sizes": [
3967
- 64,
3968
- 64
3969
- ],
3970
- "loop_orders": [
3971
- [
3972
- 0,
3973
- 1
3974
- ]
3975
- ],
3976
- "flatten_loops": [
3977
- true
3978
- ],
3979
- "l2_groupings": [
3980
- 1
3981
- ],
3982
- "range_unroll_factors": [
3983
- 0
3984
- ],
3985
- "range_warp_specializes": [],
3986
- "range_num_stages": [
3987
- 0
3988
- ],
3989
- "range_multi_buffers": [
3990
- null
3991
- ],
3992
- "range_flattens": [
3993
- null
3994
- ],
3995
- "load_eviction_policies": [
3996
- "",
3997
- "",
3998
- ""
3999
- ],
4000
- "num_warps": 8,
4001
- "num_stages": 1,
4002
- "indexing": [
4003
- "pointer",
4004
- "pointer",
4005
- "pointer",
4006
- "pointer"
4007
- ],
4008
- "pid_type": "flat"
4009
- },
4010
- "intermediate_8192_numtokens_88": {
4011
- "block_sizes": [
4012
- 64,
4013
- 128
4014
- ],
4015
- "loop_orders": [
4016
- [
4017
- 0,
4018
- 1
4019
- ]
4020
- ],
4021
- "flatten_loops": [
4022
- true
4023
- ],
4024
- "l2_groupings": [
4025
- 1
4026
- ],
4027
- "range_unroll_factors": [
4028
- 0
4029
- ],
4030
- "range_warp_specializes": [],
4031
- "range_num_stages": [
4032
- 0
4033
- ],
4034
- "range_multi_buffers": [
4035
- null
4036
- ],
4037
- "range_flattens": [
4038
- null
4039
- ],
4040
- "load_eviction_policies": [
4041
- "",
4042
- "",
4043
- ""
4044
- ],
4045
- "num_warps": 8,
4046
- "num_stages": 1,
4047
- "indexing": [
4048
- "pointer",
4049
- "pointer",
4050
- "pointer",
4051
- "pointer"
4052
- ],
4053
- "pid_type": "flat"
4054
- },
4055
- "intermediate_11008_numtokens_88": {
4056
- "block_sizes": [
4057
- 16,
4058
- 2048
4059
- ],
4060
- "loop_orders": [
4061
- [
4062
- 0,
4063
- 1
4064
- ]
4065
- ],
4066
- "flatten_loops": [
4067
- true
4068
- ],
4069
- "l2_groupings": [
4070
- 2
4071
- ],
4072
- "range_unroll_factors": [
4073
- 0
4074
- ],
4075
- "range_warp_specializes": [],
4076
- "range_num_stages": [
4077
- 0
4078
- ],
4079
- "range_multi_buffers": [
4080
- null
4081
- ],
4082
- "range_flattens": [
4083
- null
4084
- ],
4085
- "load_eviction_policies": [
4086
- "first",
4087
- "first",
4088
- ""
4089
- ],
4090
- "num_warps": 32,
4091
- "num_stages": 2,
4092
- "indexing": [
4093
- "tensor_descriptor",
4094
- "tensor_descriptor",
4095
- "pointer",
4096
- "pointer"
4097
- ],
4098
- "pid_type": "flat"
4099
- },
4100
- "intermediate_14336_numtokens_88": {
4101
- "block_sizes": [
4102
- 4,
4103
- 512
4104
- ],
4105
- "loop_orders": [
4106
- [
4107
- 0,
4108
- 1
4109
- ]
4110
- ],
4111
- "flatten_loops": [
4112
- true
4113
- ],
4114
- "l2_groupings": [
4115
- 16
4116
- ],
4117
- "range_unroll_factors": [
4118
- 0
4119
- ],
4120
- "range_warp_specializes": [],
4121
- "range_num_stages": [
4122
- 0
4123
- ],
4124
- "range_multi_buffers": [
4125
- null
4126
- ],
4127
- "range_flattens": [
4128
- null
4129
- ],
4130
- "load_eviction_policies": [
4131
- "",
4132
- "",
4133
- ""
4134
- ],
4135
- "num_warps": 32,
4136
- "num_stages": 5,
4137
- "indexing": [
4138
- "pointer",
4139
- "pointer",
4140
- "pointer",
4141
- "pointer"
4142
- ],
4143
- "pid_type": "flat"
4144
- },
4145
- "intermediate_2048_numtokens_96": {
4146
- "block_sizes": [
4147
- 64,
4148
- 32
4149
- ],
4150
- "loop_orders": [
4151
- [
4152
- 0,
4153
- 1
4154
- ]
4155
- ],
4156
- "flatten_loops": [
4157
- true
4158
- ],
4159
- "l2_groupings": [
4160
- 1
4161
- ],
4162
- "range_unroll_factors": [
4163
- 0
4164
- ],
4165
- "range_warp_specializes": [],
4166
- "range_num_stages": [
4167
- 0
4168
- ],
4169
- "range_multi_buffers": [
4170
- null
4171
- ],
4172
- "range_flattens": [
4173
- null
4174
- ],
4175
- "load_eviction_policies": [
4176
- "",
4177
- "",
4178
- ""
4179
- ],
4180
- "num_warps": 8,
4181
- "num_stages": 1,
4182
- "indexing": [
4183
- "pointer",
4184
- "pointer",
4185
- "pointer",
4186
- "pointer"
4187
- ],
4188
- "pid_type": "flat"
4189
- },
4190
- "intermediate_2880_numtokens_96": {
4191
- "block_sizes": [
4192
- 32,
4193
- 64
4194
- ],
4195
- "loop_orders": [
4196
- [
4197
- 0,
4198
- 1
4199
- ]
4200
- ],
4201
- "flatten_loops": [
4202
- true
4203
- ],
4204
- "l2_groupings": [
4205
- 1
4206
- ],
4207
- "range_unroll_factors": [
4208
- 0
4209
- ],
4210
- "range_warp_specializes": [],
4211
- "range_num_stages": [
4212
- 0
4213
- ],
4214
- "range_multi_buffers": [
4215
- null
4216
- ],
4217
- "range_flattens": [
4218
- null
4219
- ],
4220
- "load_eviction_policies": [
4221
- "",
4222
- "",
4223
- ""
4224
- ],
4225
- "num_warps": 8,
4226
- "num_stages": 1,
4227
- "indexing": [
4228
- "pointer",
4229
- "pointer",
4230
- "pointer",
4231
- "pointer"
4232
- ],
4233
- "pid_type": "flat"
4234
- },
4235
- "intermediate_4096_numtokens_96": {
4236
- "block_sizes": [
4237
- 1,
4238
- 4096
4239
- ],
4240
- "loop_orders": [
4241
- [
4242
- 1,
4243
- 0
4244
- ]
4245
- ],
4246
- "flatten_loops": [
4247
- false
4248
- ],
4249
- "l2_groupings": [
4250
- 64
4251
- ],
4252
- "range_unroll_factors": [
4253
- 0
4254
- ],
4255
- "range_warp_specializes": [],
4256
- "range_num_stages": [
4257
- 0
4258
- ],
4259
- "range_multi_buffers": [
4260
- null
4261
- ],
4262
- "range_flattens": [
4263
- null
4264
- ],
4265
- "load_eviction_policies": [
4266
- "first",
4267
- "last",
4268
- "last"
4269
- ],
4270
- "num_warps": 4,
4271
- "num_stages": 6,
4272
- "indexing": [
4273
- "pointer",
4274
- "pointer",
4275
- "tensor_descriptor",
4276
- "pointer"
4277
- ],
4278
- "pid_type": "flat"
4279
- },
4280
- "intermediate_8192_numtokens_96": {
4281
- "block_sizes": [
4282
- 64,
4283
- 128
4284
- ],
4285
- "loop_orders": [
4286
- [
4287
- 0,
4288
- 1
4289
- ]
4290
- ],
4291
- "flatten_loops": [
4292
- true
4293
- ],
4294
- "l2_groupings": [
4295
- 1
4296
- ],
4297
- "range_unroll_factors": [
4298
- 0
4299
- ],
4300
- "range_warp_specializes": [],
4301
- "range_num_stages": [
4302
- 0
4303
- ],
4304
- "range_multi_buffers": [
4305
- null
4306
- ],
4307
- "range_flattens": [
4308
- null
4309
- ],
4310
- "load_eviction_policies": [
4311
- "",
4312
- "",
4313
- ""
4314
- ],
4315
- "num_warps": 8,
4316
- "num_stages": 1,
4317
- "indexing": [
4318
- "pointer",
4319
- "pointer",
4320
- "pointer",
4321
- "pointer"
4322
- ],
4323
- "pid_type": "flat"
4324
- },
4325
- "intermediate_11008_numtokens_96": {
4326
- "block_sizes": [
4327
- 1,
4328
- 2048
4329
- ],
4330
- "loop_orders": [
4331
- [
4332
- 1,
4333
- 0
4334
- ]
4335
- ],
4336
- "flatten_loops": [
4337
- false
4338
- ],
4339
- "l2_groupings": [
4340
- 64
4341
- ],
4342
- "range_unroll_factors": [
4343
- 0
4344
- ],
4345
- "range_warp_specializes": [],
4346
- "range_num_stages": [
4347
- 0
4348
- ],
4349
- "range_multi_buffers": [
4350
- null
4351
- ],
4352
- "range_flattens": [
4353
- null
4354
- ],
4355
- "load_eviction_policies": [
4356
- "first",
4357
- "",
4358
- ""
4359
- ],
4360
- "num_warps": 32,
4361
- "num_stages": 2,
4362
- "indexing": [
4363
- "tensor_descriptor",
4364
- "pointer",
4365
- "pointer",
4366
- "tensor_descriptor"
4367
- ],
4368
- "pid_type": "flat"
4369
- },
4370
- "intermediate_14336_numtokens_96": {
4371
- "block_sizes": [
4372
- 4,
4373
- 4096
4374
- ],
4375
- "loop_orders": [
4376
- [
4377
- 0,
4378
- 1
4379
- ]
4380
- ],
4381
- "flatten_loops": [
4382
- false
4383
- ],
4384
- "l2_groupings": [
4385
- 32
4386
- ],
4387
- "range_unroll_factors": [
4388
- 0
4389
- ],
4390
- "range_warp_specializes": [],
4391
- "range_num_stages": [
4392
- 0
4393
- ],
4394
- "range_multi_buffers": [
4395
- null
4396
- ],
4397
- "range_flattens": [
4398
- null
4399
- ],
4400
- "load_eviction_policies": [
4401
- "last",
4402
- "first",
4403
- ""
4404
- ],
4405
- "num_warps": 4,
4406
- "num_stages": 1,
4407
- "indexing": [
4408
- "tensor_descriptor",
4409
- "tensor_descriptor",
4410
- "pointer",
4411
- "tensor_descriptor"
4412
- ],
4413
- "pid_type": "flat"
4414
- },
4415
- "intermediate_2048_numtokens_104": {
4416
- "block_sizes": [
4417
- 64,
4418
- 32
4419
- ],
4420
- "loop_orders": [
4421
- [
4422
- 0,
4423
- 1
4424
- ]
4425
- ],
4426
- "flatten_loops": [
4427
- true
4428
- ],
4429
- "l2_groupings": [
4430
- 1
4431
- ],
4432
- "range_unroll_factors": [
4433
- 0
4434
- ],
4435
- "range_warp_specializes": [],
4436
- "range_num_stages": [
4437
- 0
4438
- ],
4439
- "range_multi_buffers": [
4440
- null
4441
- ],
4442
- "range_flattens": [
4443
- null
4444
- ],
4445
- "load_eviction_policies": [
4446
- "",
4447
- "",
4448
- ""
4449
- ],
4450
- "num_warps": 8,
4451
- "num_stages": 1,
4452
- "indexing": [
4453
- "pointer",
4454
- "pointer",
4455
- "pointer",
4456
- "pointer"
4457
- ],
4458
- "pid_type": "flat"
4459
- },
4460
- "intermediate_2880_numtokens_104": {
4461
- "block_sizes": [
4462
- 8,
4463
- 256
4464
- ],
4465
- "loop_orders": [
4466
- [
4467
- 0,
4468
- 1
4469
- ]
4470
- ],
4471
- "flatten_loops": [
4472
- true
4473
- ],
4474
- "l2_groupings": [
4475
- 1
4476
- ],
4477
- "range_unroll_factors": [
4478
- 0
4479
- ],
4480
- "range_warp_specializes": [],
4481
- "range_num_stages": [
4482
- 0
4483
- ],
4484
- "range_multi_buffers": [
4485
- null
4486
- ],
4487
- "range_flattens": [
4488
- null
4489
- ],
4490
- "load_eviction_policies": [
4491
- "",
4492
- "",
4493
- ""
4494
- ],
4495
- "num_warps": 8,
4496
- "num_stages": 1,
4497
- "indexing": [
4498
- "pointer",
4499
- "pointer",
4500
- "pointer",
4501
- "pointer"
4502
- ],
4503
- "pid_type": "flat"
4504
- },
4505
- "intermediate_4096_numtokens_104": {
4506
- "block_sizes": [
4507
- 64,
4508
- 64
4509
- ],
4510
- "loop_orders": [
4511
- [
4512
- 0,
4513
- 1
4514
- ]
4515
- ],
4516
- "flatten_loops": [
4517
- true
4518
- ],
4519
- "l2_groupings": [
4520
- 1
4521
- ],
4522
- "range_unroll_factors": [
4523
- 0
4524
- ],
4525
- "range_warp_specializes": [],
4526
- "range_num_stages": [
4527
- 0
4528
- ],
4529
- "range_multi_buffers": [
4530
- null
4531
- ],
4532
- "range_flattens": [
4533
- null
4534
- ],
4535
- "load_eviction_policies": [
4536
- "",
4537
- "",
4538
- ""
4539
- ],
4540
- "num_warps": 8,
4541
- "num_stages": 1,
4542
- "indexing": [
4543
- "pointer",
4544
- "pointer",
4545
- "pointer",
4546
- "pointer"
4547
- ],
4548
- "pid_type": "flat"
4549
- },
4550
- "intermediate_8192_numtokens_104": {
4551
- "block_sizes": [
4552
- 64,
4553
- 128
4554
- ],
4555
- "loop_orders": [
4556
- [
4557
- 0,
4558
- 1
4559
- ]
4560
- ],
4561
- "flatten_loops": [
4562
- true
4563
- ],
4564
- "l2_groupings": [
4565
- 1
4566
- ],
4567
- "range_unroll_factors": [
4568
- 0
4569
- ],
4570
- "range_warp_specializes": [],
4571
- "range_num_stages": [
4572
- 0
4573
- ],
4574
- "range_multi_buffers": [
4575
- null
4576
- ],
4577
- "range_flattens": [
4578
- null
4579
- ],
4580
- "load_eviction_policies": [
4581
- "",
4582
- "",
4583
- ""
4584
- ],
4585
- "num_warps": 8,
4586
- "num_stages": 1,
4587
- "indexing": [
4588
- "pointer",
4589
- "pointer",
4590
- "pointer",
4591
- "pointer"
4592
- ],
4593
- "pid_type": "flat"
4594
- },
4595
- "intermediate_11008_numtokens_104": {
4596
- "block_sizes": [
4597
- 2,
4598
- 8192
4599
- ],
4600
- "loop_orders": [
4601
- [
4602
- 0,
4603
- 1
4604
- ]
4605
- ],
4606
- "flatten_loops": [
4607
- true
4608
- ],
4609
- "l2_groupings": [
4610
- 1
4611
- ],
4612
- "range_unroll_factors": [
4613
- 0
4614
- ],
4615
- "range_warp_specializes": [],
4616
- "range_num_stages": [
4617
- 0
4618
- ],
4619
- "range_multi_buffers": [
4620
- null
4621
- ],
4622
- "range_flattens": [
4623
- null
4624
- ],
4625
- "load_eviction_policies": [
4626
- "",
4627
- "last",
4628
- "first"
4629
- ],
4630
- "num_warps": 8,
4631
- "num_stages": 1,
4632
- "indexing": [
4633
- "pointer",
4634
- "tensor_descriptor",
4635
- "tensor_descriptor",
4636
- "pointer"
4637
- ],
4638
- "pid_type": "flat"
4639
- },
4640
- "intermediate_14336_numtokens_104": {
4641
- "block_sizes": [
4642
- 8,
4643
- 512
4644
- ],
4645
- "loop_orders": [
4646
- [
4647
- 0,
4648
- 1
4649
- ]
4650
- ],
4651
- "flatten_loops": [
4652
- true
4653
- ],
4654
- "l2_groupings": [
4655
- 2
4656
- ],
4657
- "range_unroll_factors": [
4658
- 0
4659
- ],
4660
- "range_warp_specializes": [],
4661
- "range_num_stages": [
4662
- 0
4663
- ],
4664
- "range_multi_buffers": [
4665
- null
4666
- ],
4667
- "range_flattens": [
4668
- null
4669
- ],
4670
- "load_eviction_policies": [
4671
- "last",
4672
- "last",
4673
- ""
4674
- ],
4675
- "num_warps": 8,
4676
- "num_stages": 1,
4677
- "indexing": [
4678
- "pointer",
4679
- "tensor_descriptor",
4680
- "pointer",
4681
- "tensor_descriptor"
4682
- ],
4683
- "pid_type": "flat"
4684
- },
4685
- "intermediate_2048_numtokens_112": {
4686
- "block_sizes": [
4687
- 64,
4688
- 32
4689
- ],
4690
- "loop_orders": [
4691
- [
4692
- 0,
4693
- 1
4694
- ]
4695
- ],
4696
- "flatten_loops": [
4697
- true
4698
- ],
4699
- "l2_groupings": [
4700
- 1
4701
- ],
4702
- "range_unroll_factors": [
4703
- 0
4704
- ],
4705
- "range_warp_specializes": [],
4706
- "range_num_stages": [
4707
- 0
4708
- ],
4709
- "range_multi_buffers": [
4710
- null
4711
- ],
4712
- "range_flattens": [
4713
- null
4714
- ],
4715
- "load_eviction_policies": [
4716
- "",
4717
- "",
4718
- ""
4719
- ],
4720
- "num_warps": 8,
4721
- "num_stages": 1,
4722
- "indexing": [
4723
- "pointer",
4724
- "pointer",
4725
- "pointer",
4726
- "pointer"
4727
- ],
4728
- "pid_type": "flat"
4729
- },
4730
- "intermediate_2880_numtokens_112": {
4731
- "block_sizes": [
4732
- 2,
4733
- 2048
4734
- ],
4735
- "loop_orders": [
4736
- [
4737
- 0,
4738
- 1
4739
- ]
4740
- ],
4741
- "flatten_loops": [
4742
- true
4743
- ],
4744
- "l2_groupings": [
4745
- 1
4746
- ],
4747
- "range_unroll_factors": [
4748
- 0
4749
- ],
4750
- "range_warp_specializes": [],
4751
- "range_num_stages": [
4752
- 0
4753
- ],
4754
- "range_multi_buffers": [
4755
- null
4756
- ],
4757
- "range_flattens": [
4758
- null
4759
- ],
4760
- "load_eviction_policies": [
4761
- "",
4762
- "",
4763
- ""
4764
- ],
4765
- "num_warps": 8,
4766
- "num_stages": 1,
4767
- "indexing": [
4768
- "pointer",
4769
- "pointer",
4770
- "pointer",
4771
- "pointer"
4772
- ],
4773
- "pid_type": "flat"
4774
- },
4775
- "intermediate_4096_numtokens_112": {
4776
- "block_sizes": [
4777
- 64,
4778
- 64
4779
- ],
4780
- "loop_orders": [
4781
- [
4782
- 0,
4783
- 1
4784
- ]
4785
- ],
4786
- "flatten_loops": [
4787
- true
4788
- ],
4789
- "l2_groupings": [
4790
- 1
4791
- ],
4792
- "range_unroll_factors": [
4793
- 0
4794
- ],
4795
- "range_warp_specializes": [],
4796
- "range_num_stages": [
4797
- 0
4798
- ],
4799
- "range_multi_buffers": [
4800
- null
4801
- ],
4802
- "range_flattens": [
4803
- null
4804
- ],
4805
- "load_eviction_policies": [
4806
- "",
4807
- "",
4808
- ""
4809
- ],
4810
- "num_warps": 8,
4811
- "num_stages": 1,
4812
- "indexing": [
4813
- "pointer",
4814
- "pointer",
4815
- "pointer",
4816
- "pointer"
4817
- ],
4818
- "pid_type": "flat"
4819
- },
4820
- "intermediate_8192_numtokens_112": {
4821
- "block_sizes": [
4822
- 4,
4823
- 512
4824
- ],
4825
- "loop_orders": [
4826
- [
4827
- 0,
4828
- 1
4829
- ]
4830
- ],
4831
- "flatten_loops": [
4832
- true
4833
- ],
4834
- "l2_groupings": [
4835
- 16
4836
- ],
4837
- "range_unroll_factors": [
4838
- 0
4839
- ],
4840
- "range_warp_specializes": [],
4841
- "range_num_stages": [
4842
- 0
4843
- ],
4844
- "range_multi_buffers": [
4845
- null
4846
- ],
4847
- "range_flattens": [
4848
- null
4849
- ],
4850
- "load_eviction_policies": [
4851
- "first",
4852
- "last",
4853
- "last"
4854
- ],
4855
- "num_warps": 16,
4856
- "num_stages": 3,
4857
- "indexing": [
4858
- "pointer",
4859
- "pointer",
4860
- "tensor_descriptor",
4861
- "pointer"
4862
- ],
4863
- "pid_type": "flat"
4864
- },
4865
- "intermediate_11008_numtokens_112": {
4866
- "block_sizes": [
4867
- 4,
4868
- 1024
4869
- ],
4870
- "loop_orders": [
4871
- [
4872
- 0,
4873
- 1
4874
- ]
4875
- ],
4876
- "flatten_loops": [
4877
- true
4878
- ],
4879
- "l2_groupings": [
4880
- 2
4881
- ],
4882
- "range_unroll_factors": [
4883
- 0
4884
- ],
4885
- "range_warp_specializes": [],
4886
- "range_num_stages": [
4887
- 0
4888
- ],
4889
- "range_multi_buffers": [
4890
- null
4891
- ],
4892
- "range_flattens": [
4893
- null
4894
- ],
4895
- "load_eviction_policies": [
4896
- "",
4897
- "last",
4898
- "last"
4899
- ],
4900
- "num_warps": 32,
4901
- "num_stages": 3,
4902
- "indexing": [
4903
- "pointer",
4904
- "pointer",
4905
- "tensor_descriptor",
4906
- "tensor_descriptor"
4907
- ],
4908
- "pid_type": "flat"
4909
- },
4910
- "intermediate_14336_numtokens_112": {
4911
- "block_sizes": [
4912
- 64,
4913
- 256
4914
- ],
4915
- "loop_orders": [
4916
- [
4917
- 0,
4918
- 1
4919
- ]
4920
- ],
4921
- "flatten_loops": [
4922
- true
4923
- ],
4924
- "l2_groupings": [
4925
- 1
4926
- ],
4927
- "range_unroll_factors": [
4928
- 0
4929
- ],
4930
- "range_warp_specializes": [],
4931
- "range_num_stages": [
4932
- 0
4933
- ],
4934
- "range_multi_buffers": [
4935
- null
4936
- ],
4937
- "range_flattens": [
4938
- null
4939
- ],
4940
- "load_eviction_policies": [
4941
- "",
4942
- "",
4943
- ""
4944
- ],
4945
- "num_warps": 8,
4946
- "num_stages": 1,
4947
- "indexing": [
4948
- "pointer",
4949
- "pointer",
4950
- "pointer",
4951
- "pointer"
4952
- ],
4953
- "pid_type": "flat"
4954
- },
4955
- "intermediate_2048_numtokens_120": {
4956
- "block_sizes": [
4957
- 8,
4958
- 256
4959
- ],
4960
- "loop_orders": [
4961
- [
4962
- 0,
4963
- 1
4964
- ]
4965
- ],
4966
- "flatten_loops": [
4967
- true
4968
- ],
4969
- "l2_groupings": [
4970
- 16
4971
- ],
4972
- "range_unroll_factors": [
4973
- 0
4974
- ],
4975
- "range_warp_specializes": [],
4976
- "range_num_stages": [
4977
- 0
4978
- ],
4979
- "range_multi_buffers": [
4980
- null
4981
- ],
4982
- "range_flattens": [
4983
- null
4984
- ],
4985
- "load_eviction_policies": [
4986
- "first",
4987
- "last",
4988
- ""
4989
- ],
4990
- "num_warps": 4,
4991
- "num_stages": 6,
4992
- "indexing": [
4993
- "tensor_descriptor",
4994
- "tensor_descriptor",
4995
- "pointer",
4996
- "tensor_descriptor"
4997
- ],
4998
- "pid_type": "flat"
4999
- },
This file is too large to show in full.
vllm/kernels/helion/configs/silu_mul_fp8/nvidia_h100.json
new
+13866
@@ -0,0 +1,13866 @@
1
+{
2
+ "intermediate_2048_numtokens_256": {
3
+ "block_sizes": [
4
+ 64,
5
+ 32
6
+ ],
7
+ "loop_orders": [
8
+ [
9
+ 0,
10
+ 1
11
+ ]
12
+ ],
13
+ "flatten_loops": [
14
+ true
15
+ ],
16
+ "l2_groupings": [
17
+ 1
18
+ ],
19
+ "range_unroll_factors": [
20
+ 0
21
+ ],
22
+ "range_warp_specializes": [],
23
+ "range_num_stages": [
24
+ 0
25
+ ],
26
+ "range_multi_buffers": [
27
+ null
28
+ ],
29
+ "range_flattens": [
30
+ null
31
+ ],
32
+ "load_eviction_policies": [
33
+ "",
34
+ "",
35
+ ""
36
+ ],
37
+ "num_warps": 8,
38
+ "num_stages": 1,
39
+ "indexing": [
40
+ "pointer",
41
+ "pointer",
42
+ "pointer",
43
+ "pointer"
44
+ ],
45
+ "pid_type": "flat"
46
+ },
47
+ "intermediate_4096_numtokens_256": {
48
+ "block_sizes": [
49
+ 256,
50
+ 32
51
+ ],
52
+ "loop_orders": [
53
+ [
54
+ 0,
55
+ 1
56
+ ]
57
+ ],
58
+ "flatten_loops": [
59
+ true
60
+ ],
61
+ "l2_groupings": [
62
+ 1
63
+ ],
64
+ "range_unroll_factors": [
65
+ 0
66
+ ],
67
+ "range_warp_specializes": [],
68
+ "range_num_stages": [
69
+ 0
70
+ ],
71
+ "range_multi_buffers": [
72
+ null
73
+ ],
74
+ "range_flattens": [
75
+ null
76
+ ],
77
+ "load_eviction_policies": [
78
+ "",
79
+ "",
80
+ ""
81
+ ],
82
+ "num_warps": 8,
83
+ "num_stages": 1,
84
+ "indexing": [
85
+ "pointer",
86
+ "pointer",
87
+ "pointer",
88
+ "pointer"
89
+ ],
90
+ "pid_type": "flat"
91
+ },
92
+ "default": {
93
+ "block_sizes": [
94
+ 1,
95
+ 512
96
+ ],
97
+ "loop_orders": [
98
+ [
99
+ 1,
100
+ 0
101
+ ]
102
+ ],
103
+ "flatten_loops": [
104
+ false
105
+ ],
106
+ "l2_groupings": [
107
+ 4
108
+ ],
109
+ "range_unroll_factors": [
110
+ 0
111
+ ],
112
+ "range_warp_specializes": [],
113
+ "range_num_stages": [
114
+ 0
115
+ ],
116
+ "range_multi_buffers": [
117
+ null
118
+ ],
119
+ "range_flattens": [
120
+ null
121
+ ],
122
+ "load_eviction_policies": [
123
+ "",
124
+ "first",
125
+ ""
126
+ ],
127
+ "num_warps": 8,
128
+ "num_stages": 2,
129
+ "indexing": [
130
+ "tensor_descriptor",
131
+ "tensor_descriptor",
132
+ "tensor_descriptor",
133
+ "pointer"
134
+ ],
135
+ "pid_type": "flat"
136
+ },
137
+ "intermediate_8192_numtokens_256": {
138
+ "block_sizes": [
139
+ 256,
140
+ 64
141
+ ],
142
+ "loop_orders": [
143
+ [
144
+ 0,
145
+ 1
146
+ ]
147
+ ],
148
+ "flatten_loops": [
149
+ true
150
+ ],
151
+ "l2_groupings": [
152
+ 1
153
+ ],
154
+ "range_unroll_factors": [
155
+ 0
156
+ ],
157
+ "range_warp_specializes": [],
158
+ "range_num_stages": [
159
+ 0
160
+ ],
161
+ "range_multi_buffers": [
162
+ null
163
+ ],
164
+ "range_flattens": [
165
+ null
166
+ ],
167
+ "load_eviction_policies": [
168
+ "",
169
+ "",
170
+ ""
171
+ ],
172
+ "num_warps": 8,
173
+ "num_stages": 1,
174
+ "indexing": [
175
+ "pointer",
176
+ "pointer",
177
+ "pointer",
178
+ "pointer"
179
+ ],
180
+ "pid_type": "flat"
181
+ },
182
+ "intermediate_14336_numtokens_256": {
183
+ "block_sizes": [
184
+ 8,
185
+ 4096
186
+ ],
187
+ "loop_orders": [
188
+ [
189
+ 0,
190
+ 1
191
+ ]
192
+ ],
193
+ "flatten_loops": [
194
+ true
195
+ ],
196
+ "l2_groupings": [
197
+ 1
198
+ ],
199
+ "range_unroll_factors": [
200
+ 0
201
+ ],
202
+ "range_warp_specializes": [],
203
+ "range_num_stages": [
204
+ 0
205
+ ],
206
+ "range_multi_buffers": [
207
+ null
208
+ ],
209
+ "range_flattens": [
210
+ null
211
+ ],
212
+ "load_eviction_policies": [
213
+ "",
214
+ "",
215
+ ""
216
+ ],
217
+ "num_warps": 8,
218
+ "num_stages": 1,
219
+ "indexing": [
220
+ "pointer",
221
+ "pointer",
222
+ "pointer",
223
+ "pointer"
224
+ ],
225
+ "pid_type": "flat"
226
+ },
227
+ "intermediate_11008_numtokens_256": {
228
+ "block_sizes": [
229
+ 64,
230
+ 32
231
+ ],
232
+ "loop_orders": [
233
+ [
234
+ 0,
235
+ 1
236
+ ]
237
+ ],
238
+ "flatten_loops": [
239
+ true
240
+ ],
241
+ "l2_groupings": [
242
+ 1
243
+ ],
244
+ "range_unroll_factors": [
245
+ 0
246
+ ],
247
+ "range_warp_specializes": [],
248
+ "range_num_stages": [
249
+ 0
250
+ ],
251
+ "range_multi_buffers": [
252
+ null
253
+ ],
254
+ "range_flattens": [
255
+ null
256
+ ],
257
+ "load_eviction_policies": [
258
+ "",
259
+ "",
260
+ ""
261
+ ],
262
+ "num_warps": 2,
263
+ "num_stages": 1,
264
+ "indexing": [
265
+ "pointer",
266
+ "pointer",
267
+ "pointer",
268
+ "pointer"
269
+ ],
270
+ "pid_type": "flat"
271
+ },
272
+ "intermediate_7688_numtokens_256": {
273
+ "block_sizes": [
274
+ 32,
275
+ 512
276
+ ],
277
+ "loop_orders": [
278
+ [
279
+ 0,
280
+ 1
281
+ ]
282
+ ],
283
+ "flatten_loops": [
284
+ true
285
+ ],
286
+ "l2_groupings": [
287
+ 1
288
+ ],
289
+ "range_unroll_factors": [
290
+ 0
291
+ ],
292
+ "range_warp_specializes": [],
293
+ "range_num_stages": [
294
+ 0
295
+ ],
296
+ "range_multi_buffers": [
297
+ null
298
+ ],
299
+ "range_flattens": [
300
+ null
301
+ ],
302
+ "load_eviction_policies": [
303
+ "",
304
+ "",
305
+ ""
306
+ ],
307
+ "num_warps": 8,
308
+ "num_stages": 1,
309
+ "indexing": [
310
+ "pointer",
311
+ "pointer",
312
+ "pointer",
313
+ "pointer"
314
+ ],
315
+ "pid_type": "flat"
316
+ },
317
+ "intermediate_2880_numtokens_256": {
318
+ "block_sizes": [
319
+ 32,
320
+ 64
321
+ ],
322
+ "loop_orders": [
323
+ [
324
+ 0,
325
+ 1
326
+ ]
327
+ ],
328
+ "flatten_loops": [
329
+ true
330
+ ],
331
+ "l2_groupings": [
332
+ 2
333
+ ],
334
+ "range_unroll_factors": [
335
+ 0
336
+ ],
337
+ "range_warp_specializes": [],
338
+ "range_num_stages": [
339
+ 0
340
+ ],
341
+ "range_multi_buffers": [
342
+ null
343
+ ],
344
+ "range_flattens": [
345
+ null
346
+ ],
347
+ "load_eviction_policies": [
348
+ "",
349
+ "",
350
+ ""
351
+ ],
352
+ "num_warps": 8,
353
+ "num_stages": 1,
354
+ "indexing": [
355
+ "pointer",
356
+ "pointer",
357
+ "pointer",
358
+ "pointer"
359
+ ],
360
+ "pid_type": "flat"
361
+ },
362
+ "intermediate_2048_numtokens_1": {
363
+ "block_sizes": [
364
+ 1,
365
+ 256
366
+ ],
367
+ "loop_orders": [
368
+ [
369
+ 0,
370
+ 1
371
+ ]
372
+ ],
373
+ "flatten_loops": [
374
+ true
375
+ ],
376
+ "l2_groupings": [
377
+ 1
378
+ ],
379
+ "range_unroll_factors": [
380
+ 0
381
+ ],
382
+ "range_warp_specializes": [],
383
+ "range_num_stages": [
384
+ 0
385
+ ],
386
+ "range_multi_buffers": [
387
+ null
388
+ ],
389
+ "range_flattens": [
390
+ null
391
+ ],
392
+ "load_eviction_policies": [
393
+ "",
394
+ "",
395
+ ""
396
+ ],
397
+ "num_warps": 2,
398
+ "num_stages": 1,
399
+ "indexing": [
400
+ "pointer",
401
+ "pointer",
402
+ "pointer",
403
+ "pointer"
404
+ ],
405
+ "pid_type": "flat"
406
+ },
407
+ "intermediate_2880_numtokens_1": {
408
+ "block_sizes": [
409
+ 1,
410
+ 128
411
+ ],
412
+ "loop_orders": [
413
+ [
414
+ 0,
415
+ 1
416
+ ]
417
+ ],
418
+ "flatten_loops": [
419
+ true
420
+ ],
421
+ "l2_groupings": [
422
+ 1
423
+ ],
424
+ "range_unroll_factors": [
425
+ 0
426
+ ],
427
+ "range_warp_specializes": [],
428
+ "range_num_stages": [
429
+ 0
430
+ ],
431
+ "range_multi_buffers": [
432
+ null
433
+ ],
434
+ "range_flattens": [
435
+ null
436
+ ],
437
+ "load_eviction_policies": [
438
+ "",
439
+ "",
440
+ ""
441
+ ],
442
+ "num_warps": 4,
443
+ "num_stages": 1,
444
+ "indexing": [
445
+ "pointer",
446
+ "pointer",
447
+ "pointer",
448
+ "pointer"
449
+ ],
450
+ "pid_type": "flat"
451
+ },
452
+ "intermediate_4096_numtokens_1": {
453
+ "block_sizes": [
454
+ 1,
455
+ 256
456
+ ],
457
+ "loop_orders": [
458
+ [
459
+ 0,
460
+ 1
461
+ ]
462
+ ],
463
+ "flatten_loops": [
464
+ true
465
+ ],
466
+ "l2_groupings": [
467
+ 1
468
+ ],
469
+ "range_unroll_factors": [
470
+ 0
471
+ ],
472
+ "range_warp_specializes": [],
473
+ "range_num_stages": [
474
+ 0
475
+ ],
476
+ "range_multi_buffers": [
477
+ null
478
+ ],
479
+ "range_flattens": [
480
+ null
481
+ ],
482
+ "load_eviction_policies": [
483
+ "",
484
+ "",
485
+ ""
486
+ ],
487
+ "num_warps": 2,
488
+ "num_stages": 1,
489
+ "indexing": [
490
+ "pointer",
491
+ "pointer",
492
+ "pointer",
493
+ "pointer"
494
+ ],
495
+ "pid_type": "flat"
496
+ },
497
+ "intermediate_8192_numtokens_1": {
498
+ "block_sizes": [
499
+ 1,
500
+ 256
501
+ ],
502
+ "loop_orders": [
503
+ [
504
+ 0,
505
+ 1
506
+ ]
507
+ ],
508
+ "flatten_loops": [
509
+ true
510
+ ],
511
+ "l2_groupings": [
512
+ 1
513
+ ],
514
+ "range_unroll_factors": [
515
+ 0
516
+ ],
517
+ "range_warp_specializes": [],
518
+ "range_num_stages": [
519
+ 0
520
+ ],
521
+ "range_multi_buffers": [
522
+ null
523
+ ],
524
+ "range_flattens": [
525
+ null
526
+ ],
527
+ "load_eviction_policies": [
528
+ "",
529
+ "",
530
+ ""
531
+ ],
532
+ "num_warps": 2,
533
+ "num_stages": 1,
534
+ "indexing": [
535
+ "pointer",
536
+ "pointer",
537
+ "pointer",
538
+ "pointer"
539
+ ],
540
+ "pid_type": "flat"
541
+ },
542
+ "intermediate_11008_numtokens_1": {
543
+ "block_sizes": [
544
+ 1,
545
+ 256
546
+ ],
547
+ "loop_orders": [
548
+ [
549
+ 0,
550
+ 1
551
+ ]
552
+ ],
553
+ "flatten_loops": [
554
+ true
555
+ ],
556
+ "l2_groupings": [
557
+ 1
558
+ ],
559
+ "range_unroll_factors": [
560
+ 0
561
+ ],
562
+ "range_warp_specializes": [],
563
+ "range_num_stages": [
564
+ 0
565
+ ],
566
+ "range_multi_buffers": [
567
+ null
568
+ ],
569
+ "range_flattens": [
570
+ null
571
+ ],
572
+ "load_eviction_policies": [
573
+ "",
574
+ "",
575
+ ""
576
+ ],
577
+ "num_warps": 2,
578
+ "num_stages": 1,
579
+ "indexing": [
580
+ "pointer",
581
+ "pointer",
582
+ "pointer",
583
+ "pointer"
584
+ ],
585
+ "pid_type": "flat"
586
+ },
587
+ "intermediate_14336_numtokens_1": {
588
+ "block_sizes": [
589
+ 1,
590
+ 256
591
+ ],
592
+ "loop_orders": [
593
+ [
594
+ 0,
595
+ 1
596
+ ]
597
+ ],
598
+ "flatten_loops": [
599
+ true
600
+ ],
601
+ "l2_groupings": [
602
+ 1
603
+ ],
604
+ "range_unroll_factors": [
605
+ 0
606
+ ],
607
+ "range_warp_specializes": [],
608
+ "range_num_stages": [
609
+ 0
610
+ ],
611
+ "range_multi_buffers": [
612
+ null
613
+ ],
614
+ "range_flattens": [
615
+ null
616
+ ],
617
+ "load_eviction_policies": [
618
+ "",
619
+ "",
620
+ ""
621
+ ],
622
+ "num_warps": 2,
623
+ "num_stages": 1,
624
+ "indexing": [
625
+ "pointer",
626
+ "pointer",
627
+ "pointer",
628
+ "pointer"
629
+ ],
630
+ "pid_type": "flat"
631
+ },
632
+ "intermediate_2048_numtokens_2": {
633
+ "block_sizes": [
634
+ 2,
635
+ 128
636
+ ],
637
+ "loop_orders": [
638
+ [
639
+ 0,
640
+ 1
641
+ ]
642
+ ],
643
+ "flatten_loops": [
644
+ true
645
+ ],
646
+ "l2_groupings": [
647
+ 1
648
+ ],
649
+ "range_unroll_factors": [
650
+ 0
651
+ ],
652
+ "range_warp_specializes": [],
653
+ "range_num_stages": [
654
+ 0
655
+ ],
656
+ "range_multi_buffers": [
657
+ null
658
+ ],
659
+ "range_flattens": [
660
+ null
661
+ ],
662
+ "load_eviction_policies": [
663
+ "",
664
+ "",
665
+ ""
666
+ ],
667
+ "num_warps": 2,
668
+ "num_stages": 1,
669
+ "indexing": [
670
+ "pointer",
671
+ "pointer",
672
+ "pointer",
673
+ "pointer"
674
+ ],
675
+ "pid_type": "flat"
676
+ },
677
+ "intermediate_2880_numtokens_2": {
678
+ "block_sizes": [
679
+ 2,
680
+ 256
681
+ ],
682
+ "loop_orders": [
683
+ [
684
+ 0,
685
+ 1
686
+ ]
687
+ ],
688
+ "flatten_loops": [
689
+ true
690
+ ],
691
+ "l2_groupings": [
692
+ 1
693
+ ],
694
+ "range_unroll_factors": [
695
+ 0
696
+ ],
697
+ "range_warp_specializes": [],
698
+ "range_num_stages": [
699
+ 0
700
+ ],
701
+ "range_multi_buffers": [
702
+ null
703
+ ],
704
+ "range_flattens": [
705
+ null
706
+ ],
707
+ "load_eviction_policies": [
708
+ "",
709
+ "",
710
+ ""
711
+ ],
712
+ "num_warps": 4,
713
+ "num_stages": 1,
714
+ "indexing": [
715
+ "pointer",
716
+ "pointer",
717
+ "pointer",
718
+ "pointer"
719
+ ],
720
+ "pid_type": "flat"
721
+ },
722
+ "intermediate_4096_numtokens_2": {
723
+ "block_sizes": [
724
+ 2,
725
+ 128
726
+ ],
727
+ "loop_orders": [
728
+ [
729
+ 0,
730
+ 1
731
+ ]
732
+ ],
733
+ "flatten_loops": [
734
+ true
735
+ ],
736
+ "l2_groupings": [
737
+ 1
738
+ ],
739
+ "range_unroll_factors": [
740
+ 0
741
+ ],
742
+ "range_warp_specializes": [],
743
+ "range_num_stages": [
744
+ 0
745
+ ],
746
+ "range_multi_buffers": [
747
+ null
748
+ ],
749
+ "range_flattens": [
750
+ null
751
+ ],
752
+ "load_eviction_policies": [
753
+ "",
754
+ "",
755
+ ""
756
+ ],
757
+ "num_warps": 2,
758
+ "num_stages": 1,
759
+ "indexing": [
760
+ "pointer",
761
+ "pointer",
762
+ "pointer",
763
+ "pointer"
764
+ ],
765
+ "pid_type": "flat"
766
+ },
767
+ "intermediate_8192_numtokens_2": {
768
+ "block_sizes": [
769
+ 2,
770
+ 128
771
+ ],
772
+ "loop_orders": [
773
+ [
774
+ 0,
775
+ 1
776
+ ]
777
+ ],
778
+ "flatten_loops": [
779
+ true
780
+ ],
781
+ "l2_groupings": [
782
+ 1
783
+ ],
784
+ "range_unroll_factors": [
785
+ 0
786
+ ],
787
+ "range_warp_specializes": [],
788
+ "range_num_stages": [
789
+ 0
790
+ ],
791
+ "range_multi_buffers": [
792
+ null
793
+ ],
794
+ "range_flattens": [
795
+ null
796
+ ],
797
+ "load_eviction_policies": [
798
+ "",
799
+ "",
800
+ ""
801
+ ],
802
+ "num_warps": 2,
803
+ "num_stages": 1,
804
+ "indexing": [
805
+ "pointer",
806
+ "pointer",
807
+ "pointer",
808
+ "pointer"
809
+ ],
810
+ "pid_type": "flat"
811
+ },
812
+ "intermediate_11008_numtokens_2": {
813
+ "block_sizes": [
814
+ 1,
815
+ 16384
816
+ ],
817
+ "loop_orders": [
818
+ [
819
+ 1,
820
+ 0
821
+ ]
822
+ ],
823
+ "flatten_loops": [
824
+ true
825
+ ],
826
+ "l2_groupings": [
827
+ 16
828
+ ],
829
+ "range_unroll_factors": [
830
+ 0
831
+ ],
832
+ "range_warp_specializes": [],
833
+ "range_num_stages": [
834
+ 0
835
+ ],
836
+ "range_multi_buffers": [
837
+ null
838
+ ],
839
+ "range_flattens": [
840
+ null
841
+ ],
842
+ "load_eviction_policies": [
843
+ "last",
844
+ "first",
845
+ "last"
846
+ ],
847
+ "num_warps": 32,
848
+ "num_stages": 3,
849
+ "indexing": [
850
+ "tensor_descriptor",
851
+ "pointer",
852
+ "pointer",
853
+ "tensor_descriptor"
854
+ ],
855
+ "pid_type": "xyz"
856
+ },
857
+ "intermediate_14336_numtokens_2": {
858
+ "block_sizes": [
859
+ 2,
860
+ 128
861
+ ],
862
+ "loop_orders": [
863
+ [
864
+ 0,
865
+ 1
866
+ ]
867
+ ],
868
+ "flatten_loops": [
869
+ true
870
+ ],
871
+ "l2_groupings": [
872
+ 1
873
+ ],
874
+ "range_unroll_factors": [
875
+ 0
876
+ ],
877
+ "range_warp_specializes": [],
878
+ "range_num_stages": [
879
+ 0
880
+ ],
881
+ "range_multi_buffers": [
882
+ null
883
+ ],
884
+ "range_flattens": [
885
+ null
886
+ ],
887
+ "load_eviction_policies": [
888
+ "",
889
+ "",
890
+ ""
891
+ ],
892
+ "num_warps": 2,
893
+ "num_stages": 1,
894
+ "indexing": [
895
+ "pointer",
896
+ "pointer",
897
+ "pointer",
898
+ "pointer"
899
+ ],
900
+ "pid_type": "flat"
901
+ },
902
+ "intermediate_2048_numtokens_4": {
903
+ "block_sizes": [
904
+ 4,
905
+ 64
906
+ ],
907
+ "loop_orders": [
908
+ [
909
+ 0,
910
+ 1
911
+ ]
912
+ ],
913
+ "flatten_loops": [
914
+ true
915
+ ],
916
+ "l2_groupings": [
917
+ 1
918
+ ],
919
+ "range_unroll_factors": [
920
+ 0
921
+ ],
922
+ "range_warp_specializes": [],
923
+ "range_num_stages": [
924
+ 0
925
+ ],
926
+ "range_multi_buffers": [
927
+ null
928
+ ],
929
+ "range_flattens": [
930
+ null
931
+ ],
932
+ "load_eviction_policies": [
933
+ "",
934
+ "",
935
+ ""
936
+ ],
937
+ "num_warps": 2,
938
+ "num_stages": 1,
939
+ "indexing": [
940
+ "pointer",
941
+ "pointer",
942
+ "pointer",
943
+ "pointer"
944
+ ],
945
+ "pid_type": "flat"
946
+ },
947
+ "intermediate_2880_numtokens_4": {
948
+ "block_sizes": [
949
+ 4,
950
+ 64
951
+ ],
952
+ "loop_orders": [
953
+ [
954
+ 0,
955
+ 1
956
+ ]
957
+ ],
958
+ "flatten_loops": [
959
+ true
960
+ ],
961
+ "l2_groupings": [
962
+ 64
963
+ ],
964
+ "range_unroll_factors": [
965
+ 0
966
+ ],
967
+ "range_warp_specializes": [],
968
+ "range_num_stages": [
969
+ 0
970
+ ],
971
+ "range_multi_buffers": [
972
+ null
973
+ ],
974
+ "range_flattens": [
975
+ null
976
+ ],
977
+ "load_eviction_policies": [
978
+ "",
979
+ "last",
980
+ "first"
981
+ ],
982
+ "num_warps": 8,
983
+ "num_stages": 7,
984
+ "indexing": [
985
+ "pointer",
986
+ "tensor_descriptor",
987
+ "tensor_descriptor",
988
+ "pointer"
989
+ ],
990
+ "pid_type": "xyz"
991
+ },
992
+ "intermediate_4096_numtokens_4": {
993
+ "block_sizes": [
994
+ 4,
995
+ 64
996
+ ],
997
+ "loop_orders": [
998
+ [
999
+ 0,
1000
+ 1
1001
+ ]
1002
+ ],
1003
+ "flatten_loops": [
1004
+ true
1005
+ ],
1006
+ "l2_groupings": [
1007
+ 1
1008
+ ],
1009
+ "range_unroll_factors": [
1010
+ 0
1011
+ ],
1012
+ "range_warp_specializes": [],
1013
+ "range_num_stages": [
1014
+ 0
1015
+ ],
1016
+ "range_multi_buffers": [
1017
+ null
1018
+ ],
1019
+ "range_flattens": [
1020
+ null
1021
+ ],
1022
+ "load_eviction_policies": [
1023
+ "",
1024
+ "",
1025
+ ""
1026
+ ],
1027
+ "num_warps": 2,
1028
+ "num_stages": 1,
1029
+ "indexing": [
1030
+ "pointer",
1031
+ "pointer",
1032
+ "pointer",
1033
+ "pointer"
1034
+ ],
1035
+ "pid_type": "flat"
1036
+ },
1037
+ "intermediate_8192_numtokens_4": {
1038
+ "block_sizes": [
1039
+ 1,
1040
+ 2048
1041
+ ],
1042
+ "loop_orders": [
1043
+ [
1044
+ 0,
1045
+ 1
1046
+ ]
1047
+ ],
1048
+ "flatten_loops": [
1049
+ false
1050
+ ],
1051
+ "l2_groupings": [
1052
+ 1
1053
+ ],
1054
+ "range_unroll_factors": [
1055
+ 0
1056
+ ],
1057
+ "range_warp_specializes": [],
1058
+ "range_num_stages": [
1059
+ 0
1060
+ ],
1061
+ "range_multi_buffers": [
1062
+ null
1063
+ ],
1064
+ "range_flattens": [
1065
+ null
1066
+ ],
1067
+ "load_eviction_policies": [
1068
+ "",
1069
+ "",
1070
+ ""
1071
+ ],
1072
+ "num_warps": 8,
1073
+ "num_stages": 8,
1074
+ "indexing": [
1075
+ "pointer",
1076
+ "tensor_descriptor",
1077
+ "pointer",
1078
+ "pointer"
1079
+ ],
1080
+ "pid_type": "flat"
1081
+ },
1082
+ "intermediate_11008_numtokens_4": {
1083
+ "block_sizes": [
1084
+ 4,
1085
+ 2048
1086
+ ],
1087
+ "loop_orders": [
1088
+ [
1089
+ 1,
1090
+ 0
1091
+ ]
1092
+ ],
1093
+ "flatten_loops": [
1094
+ true
1095
+ ],
1096
+ "l2_groupings": [
1097
+ 4
1098
+ ],
1099
+ "range_unroll_factors": [
1100
+ 0
1101
+ ],
1102
+ "range_warp_specializes": [],
1103
+ "range_num_stages": [
1104
+ 0
1105
+ ],
1106
+ "range_multi_buffers": [
1107
+ null
1108
+ ],
1109
+ "range_flattens": [
1110
+ null
1111
+ ],
1112
+ "load_eviction_policies": [
1113
+ "first",
1114
+ "last",
1115
+ "last"
1116
+ ],
1117
+ "num_warps": 8,
1118
+ "num_stages": 6,
1119
+ "indexing": [
1120
+ "pointer",
1121
+ "tensor_descriptor",
1122
+ "tensor_descriptor",
1123
+ "tensor_descriptor"
1124
+ ],
1125
+ "pid_type": "xyz"
1126
+ },
1127
+ "intermediate_14336_numtokens_4": {
1128
+ "block_sizes": [
1129
+ 4,
1130
+ 256
1131
+ ],
1132
+ "loop_orders": [
1133
+ [
1134
+ 0,
1135
+ 1
1136
+ ]
1137
+ ],
1138
+ "flatten_loops": [
1139
+ true
1140
+ ],
1141
+ "l2_groupings": [
1142
+ 64
1143
+ ],
1144
+ "range_unroll_factors": [
1145
+ 0
1146
+ ],
1147
+ "range_warp_specializes": [],
1148
+ "range_num_stages": [
1149
+ 0
1150
+ ],
1151
+ "range_multi_buffers": [
1152
+ null
1153
+ ],
1154
+ "range_flattens": [
1155
+ null
1156
+ ],
1157
+ "load_eviction_policies": [
1158
+ "",
1159
+ "",
1160
+ ""
1161
+ ],
1162
+ "num_warps": 2,
1163
+ "num_stages": 2,
1164
+ "indexing": [
1165
+ "pointer",
1166
+ "pointer",
1167
+ "pointer",
1168
+ "pointer"
1169
+ ],
1170
+ "pid_type": "xyz"
1171
+ },
1172
+ "intermediate_2048_numtokens_8": {
1173
+ "block_sizes": [
1174
+ 8,
1175
+ 32
1176
+ ],
1177
+ "loop_orders": [
1178
+ [
1179
+ 0,
1180
+ 1
1181
+ ]
1182
+ ],
1183
+ "flatten_loops": [
1184
+ true
1185
+ ],
1186
+ "l2_groupings": [
1187
+ 1
1188
+ ],
1189
+ "range_unroll_factors": [
1190
+ 0
1191
+ ],
1192
+ "range_warp_specializes": [],
1193
+ "range_num_stages": [
1194
+ 0
1195
+ ],
1196
+ "range_multi_buffers": [
1197
+ null
1198
+ ],
1199
+ "range_flattens": [
1200
+ null
1201
+ ],
1202
+ "load_eviction_policies": [
1203
+ "",
1204
+ "",
1205
+ ""
1206
+ ],
1207
+ "num_warps": 2,
1208
+ "num_stages": 1,
1209
+ "indexing": [
1210
+ "pointer",
1211
+ "pointer",
1212
+ "pointer",
1213
+ "pointer"
1214
+ ],
1215
+ "pid_type": "flat"
1216
+ },
1217
+ "intermediate_2880_numtokens_8": {
1218
+ "block_sizes": [
1219
+ 4,
1220
+ 2048
1221
+ ],
1222
+ "loop_orders": [
1223
+ [
1224
+ 0,
1225
+ 1
1226
+ ]
1227
+ ],
1228
+ "flatten_loops": [
1229
+ true
1230
+ ],
1231
+ "l2_groupings": [
1232
+ 4
1233
+ ],
1234
+ "range_unroll_factors": [
1235
+ 0
1236
+ ],
1237
+ "range_warp_specializes": [],
1238
+ "range_num_stages": [
1239
+ 0
1240
+ ],
1241
+ "range_multi_buffers": [
1242
+ null
1243
+ ],
1244
+ "range_flattens": [
1245
+ null
1246
+ ],
1247
+ "load_eviction_policies": [
1248
+ "last",
1249
+ "last",
1250
+ ""
1251
+ ],
1252
+ "num_warps": 16,
1253
+ "num_stages": 5,
1254
+ "indexing": [
1255
+ "pointer",
1256
+ "tensor_descriptor",
1257
+ "tensor_descriptor",
1258
+ "pointer"
1259
+ ],
1260
+ "pid_type": "xyz"
1261
+ },
1262
+ "intermediate_4096_numtokens_8": {
1263
+ "block_sizes": [
1264
+ 8,
1265
+ 32
1266
+ ],
1267
+ "loop_orders": [
1268
+ [
1269
+ 0,
1270
+ 1
1271
+ ]
1272
+ ],
1273
+ "flatten_loops": [
1274
+ true
1275
+ ],
1276
+ "l2_groupings": [
1277
+ 1
1278
+ ],
1279
+ "range_unroll_factors": [
1280
+ 0
1281
+ ],
1282
+ "range_warp_specializes": [],
1283
+ "range_num_stages": [
1284
+ 0
1285
+ ],
1286
+ "range_multi_buffers": [
1287
+ null
1288
+ ],
1289
+ "range_flattens": [
1290
+ null
1291
+ ],
1292
+ "load_eviction_policies": [
1293
+ "",
1294
+ "",
1295
+ ""
1296
+ ],
1297
+ "num_warps": 2,
1298
+ "num_stages": 1,
1299
+ "indexing": [
1300
+ "pointer",
1301
+ "pointer",
1302
+ "pointer",
1303
+ "pointer"
1304
+ ],
1305
+ "pid_type": "flat"
1306
+ },
1307
+ "intermediate_8192_numtokens_8": {
1308
+ "block_sizes": [
1309
+ 2,
1310
+ 1024
1311
+ ],
1312
+ "loop_orders": [
1313
+ [
1314
+ 0,
1315
+ 1
1316
+ ]
1317
+ ],
1318
+ "flatten_loops": [
1319
+ true
1320
+ ],
1321
+ "l2_groupings": [
1322
+ 16
1323
+ ],
1324
+ "range_unroll_factors": [
1325
+ 0
1326
+ ],
1327
+ "range_warp_specializes": [],
1328
+ "range_num_stages": [
1329
+ 0
1330
+ ],
1331
+ "range_multi_buffers": [
1332
+ null
1333
+ ],
1334
+ "range_flattens": [
1335
+ null
1336
+ ],
1337
+ "load_eviction_policies": [
1338
+ "first",
1339
+ "last",
1340
+ ""
1341
+ ],
1342
+ "num_warps": 1,
1343
+ "num_stages": 8,
1344
+ "indexing": [
1345
+ "tensor_descriptor",
1346
+ "pointer",
1347
+ "tensor_descriptor",
1348
+ "pointer"
1349
+ ],
1350
+ "pid_type": "flat"
1351
+ },
1352
+ "intermediate_11008_numtokens_8": {
1353
+ "block_sizes": [
1354
+ 1,
1355
+ 1024
1356
+ ],
1357
+ "loop_orders": [
1358
+ [
1359
+ 1,
1360
+ 0
1361
+ ]
1362
+ ],
1363
+ "flatten_loops": [
1364
+ true
1365
+ ],
1366
+ "l2_groupings": [
1367
+ 4
1368
+ ],
1369
+ "range_unroll_factors": [
1370
+ 0
1371
+ ],
1372
+ "range_warp_specializes": [],
1373
+ "range_num_stages": [
1374
+ 0
1375
+ ],
1376
+ "range_multi_buffers": [
1377
+ null
1378
+ ],
1379
+ "range_flattens": [
1380
+ null
1381
+ ],
1382
+ "load_eviction_policies": [
1383
+ "last",
1384
+ "",
1385
+ "first"
1386
+ ],
1387
+ "num_warps": 2,
1388
+ "num_stages": 5,
1389
+ "indexing": [
1390
+ "tensor_descriptor",
1391
+ "tensor_descriptor",
1392
+ "tensor_descriptor",
1393
+ "tensor_descriptor"
1394
+ ],
1395
+ "pid_type": "flat"
1396
+ },
1397
+ "intermediate_14336_numtokens_8": {
1398
+ "block_sizes": [
1399
+ 8,
1400
+ 128
1401
+ ],
1402
+ "loop_orders": [
1403
+ [
1404
+ 0,
1405
+ 1
1406
+ ]
1407
+ ],
1408
+ "flatten_loops": [
1409
+ true
1410
+ ],
1411
+ "l2_groupings": [
1412
+ 1
1413
+ ],
1414
+ "range_unroll_factors": [
1415
+ 0
1416
+ ],
1417
+ "range_warp_specializes": [],
1418
+ "range_num_stages": [
1419
+ 0
1420
+ ],
1421
+ "range_multi_buffers": [
1422
+ null
1423
+ ],
1424
+ "range_flattens": [
1425
+ null
1426
+ ],
1427
+ "load_eviction_policies": [
1428
+ "",
1429
+ "",
1430
+ ""
1431
+ ],
1432
+ "num_warps": 4,
1433
+ "num_stages": 1,
1434
+ "indexing": [
1435
+ "pointer",
1436
+ "pointer",
1437
+ "pointer",
1438
+ "pointer"
1439
+ ],
1440
+ "pid_type": "flat"
1441
+ },
1442
+ "intermediate_2048_numtokens_16": {
1443
+ "block_sizes": [
1444
+ 8,
1445
+ 512
1446
+ ],
1447
+ "loop_orders": [
1448
+ [
1449
+ 0,
1450
+ 1
1451
+ ]
1452
+ ],
1453
+ "flatten_loops": [
1454
+ true
1455
+ ],
1456
+ "l2_groupings": [
1457
+ 64
1458
+ ],
1459
+ "range_unroll_factors": [
1460
+ 0
1461
+ ],
1462
+ "range_warp_specializes": [],
1463
+ "range_num_stages": [
1464
+ 0
1465
+ ],
1466
+ "range_multi_buffers": [
1467
+ null
1468
+ ],
1469
+ "range_flattens": [
1470
+ null
1471
+ ],
1472
+ "load_eviction_policies": [
1473
+ "last",
1474
+ "last",
1475
+ "first"
1476
+ ],
1477
+ "num_warps": 16,
1478
+ "num_stages": 2,
1479
+ "indexing": [
1480
+ "pointer",
1481
+ "pointer",
1482
+ "tensor_descriptor",
1483
+ "pointer"
1484
+ ],
1485
+ "pid_type": "xyz"
1486
+ },
1487
+ "intermediate_2880_numtokens_16": {
1488
+ "block_sizes": [
1489
+ 2,
1490
+ 256
1491
+ ],
1492
+ "loop_orders": [
1493
+ [
1494
+ 0,
1495
+ 1
1496
+ ]
1497
+ ],
1498
+ "flatten_loops": [
1499
+ true
1500
+ ],
1501
+ "l2_groupings": [
1502
+ 1
1503
+ ],
1504
+ "range_unroll_factors": [
1505
+ 0
1506
+ ],
1507
+ "range_warp_specializes": [],
1508
+ "range_num_stages": [
1509
+ 0
1510
+ ],
1511
+ "range_multi_buffers": [
1512
+ null
1513
+ ],
1514
+ "range_flattens": [
1515
+ null
1516
+ ],
1517
+ "load_eviction_policies": [
1518
+ "",
1519
+ "",
1520
+ ""
1521
+ ],
1522
+ "num_warps": 4,
1523
+ "num_stages": 1,
1524
+ "indexing": [
1525
+ "pointer",
1526
+ "pointer",
1527
+ "pointer",
1528
+ "pointer"
1529
+ ],
1530
+ "pid_type": "flat"
1531
+ },
1532
+ "intermediate_4096_numtokens_16": {
1533
+ "block_sizes": [
1534
+ 16,
1535
+ 256
1536
+ ],
1537
+ "loop_orders": [
1538
+ [
1539
+ 0,
1540
+ 1
1541
+ ]
1542
+ ],
1543
+ "flatten_loops": [
1544
+ true
1545
+ ],
1546
+ "l2_groupings": [
1547
+ 16
1548
+ ],
1549
+ "range_unroll_factors": [
1550
+ 0
1551
+ ],
1552
+ "range_warp_specializes": [],
1553
+ "range_num_stages": [
1554
+ 0
1555
+ ],
1556
+ "range_multi_buffers": [
1557
+ null
1558
+ ],
1559
+ "range_flattens": [
1560
+ null
1561
+ ],
1562
+ "load_eviction_policies": [
1563
+ "last",
1564
+ "",
1565
+ "last"
1566
+ ],
1567
+ "num_warps": 32,
1568
+ "num_stages": 1,
1569
+ "indexing": [
1570
+ "pointer",
1571
+ "pointer",
1572
+ "pointer",
1573
+ "pointer"
1574
+ ],
1575
+ "pid_type": "flat"
1576
+ },
1577
+ "intermediate_8192_numtokens_16": {
1578
+ "block_sizes": [
1579
+ 16,
1580
+ 64
1581
+ ],
1582
+ "loop_orders": [
1583
+ [
1584
+ 0,
1585
+ 1
1586
+ ]
1587
+ ],
1588
+ "flatten_loops": [
1589
+ true
1590
+ ],
1591
+ "l2_groupings": [
1592
+ 1
1593
+ ],
1594
+ "range_unroll_factors": [
1595
+ 0
1596
+ ],
1597
+ "range_warp_specializes": [],
1598
+ "range_num_stages": [
1599
+ 0
1600
+ ],
1601
+ "range_multi_buffers": [
1602
+ null
1603
+ ],
1604
+ "range_flattens": [
1605
+ null
1606
+ ],
1607
+ "load_eviction_policies": [
1608
+ "",
1609
+ "",
1610
+ ""
1611
+ ],
1612
+ "num_warps": 4,
1613
+ "num_stages": 1,
1614
+ "indexing": [
1615
+ "pointer",
1616
+ "pointer",
1617
+ "pointer",
1618
+ "pointer"
1619
+ ],
1620
+ "pid_type": "flat"
1621
+ },
1622
+ "intermediate_11008_numtokens_16": {
1623
+ "block_sizes": [
1624
+ 1,
1625
+ 2048
1626
+ ],
1627
+ "loop_orders": [
1628
+ [
1629
+ 0,
1630
+ 1
1631
+ ]
1632
+ ],
1633
+ "flatten_loops": [
1634
+ true
1635
+ ],
1636
+ "l2_groupings": [
1637
+ 1
1638
+ ],
1639
+ "range_unroll_factors": [
1640
+ 0
1641
+ ],
1642
+ "range_warp_specializes": [],
1643
+ "range_num_stages": [
1644
+ 0
1645
+ ],
1646
+ "range_multi_buffers": [
1647
+ null
1648
+ ],
1649
+ "range_flattens": [
1650
+ null
1651
+ ],
1652
+ "load_eviction_policies": [
1653
+ "",
1654
+ "",
1655
+ ""
1656
+ ],
1657
+ "num_warps": 8,
1658
+ "num_stages": 1,
1659
+ "indexing": [
1660
+ "pointer",
1661
+ "pointer",
1662
+ "pointer",
1663
+ "pointer"
1664
+ ],
1665
+ "pid_type": "flat"
1666
+ },
1667
+ "intermediate_14336_numtokens_16": {
1668
+ "block_sizes": [
1669
+ 2,
1670
+ 256
1671
+ ],
1672
+ "loop_orders": [
1673
+ [
1674
+ 1,
1675
+ 0
1676
+ ]
1677
+ ],
1678
+ "flatten_loops": [
1679
+ true
1680
+ ],
1681
+ "l2_groupings": [
1682
+ 64
1683
+ ],
1684
+ "range_unroll_factors": [
1685
+ 0
1686
+ ],
1687
+ "range_warp_specializes": [],
1688
+ "range_num_stages": [
1689
+ 0
1690
+ ],
1691
+ "range_multi_buffers": [
1692
+ null
1693
+ ],
1694
+ "range_flattens": [
1695
+ null
1696
+ ],
1697
+ "load_eviction_policies": [
1698
+ "last",
1699
+ "last",
1700
+ "last"
1701
+ ],
1702
+ "num_warps": 16,
1703
+ "num_stages": 7,
1704
+ "indexing": [
1705
+ "tensor_descriptor",
1706
+ "pointer",
1707
+ "tensor_descriptor",
1708
+ "tensor_descriptor"
1709
+ ],
1710
+ "pid_type": "flat"
1711
+ },
1712
+ "intermediate_2048_numtokens_24": {
1713
+ "block_sizes": [
1714
+ 1,
1715
+ 1024
1716
+ ],
1717
+ "loop_orders": [
1718
+ [
1719
+ 0,
1720
+ 1
1721
+ ]
1722
+ ],
1723
+ "flatten_loops": [
1724
+ true
1725
+ ],
1726
+ "l2_groupings": [
1727
+ 64
1728
+ ],
1729
+ "range_unroll_factors": [
1730
+ 0
1731
+ ],
1732
+ "range_warp_specializes": [],
1733
+ "range_num_stages": [
1734
+ 0
1735
+ ],
1736
+ "range_multi_buffers": [
1737
+ null
1738
+ ],
1739
+ "range_flattens": [
1740
+ null
1741
+ ],
1742
+ "load_eviction_policies": [
1743
+ "",
1744
+ "",
1745
+ "first"
1746
+ ],
1747
+ "num_warps": 4,
1748
+ "num_stages": 8,
1749
+ "indexing": [
1750
+ "tensor_descriptor",
1751
+ "pointer",
1752
+ "pointer",
1753
+ "tensor_descriptor"
1754
+ ],
1755
+ "pid_type": "flat"
1756
+ },
1757
+ "intermediate_2880_numtokens_24": {
1758
+ "block_sizes": [
1759
+ 4,
1760
+ 1024
1761
+ ],
1762
+ "loop_orders": [
1763
+ [
1764
+ 0,
1765
+ 1
1766
+ ]
1767
+ ],
1768
+ "flatten_loops": [
1769
+ true
1770
+ ],
1771
+ "l2_groupings": [
1772
+ 16
1773
+ ],
1774
+ "range_unroll_factors": [
1775
+ 0
1776
+ ],
1777
+ "range_warp_specializes": [],
1778
+ "range_num_stages": [
1779
+ 0
1780
+ ],
1781
+ "range_multi_buffers": [
1782
+ null
1783
+ ],
1784
+ "range_flattens": [
1785
+ null
1786
+ ],
1787
+ "load_eviction_policies": [
1788
+ "first",
1789
+ "",
1790
+ "first"
1791
+ ],
1792
+ "num_warps": 16,
1793
+ "num_stages": 3,
1794
+ "indexing": [
1795
+ "tensor_descriptor",
1796
+ "pointer",
1797
+ "tensor_descriptor",
1798
+ "pointer"
1799
+ ],
1800
+ "pid_type": "flat"
1801
+ },
1802
+ "intermediate_4096_numtokens_24": {
1803
+ "block_sizes": [
1804
+ 16,
1805
+ 64
1806
+ ],
1807
+ "loop_orders": [
1808
+ [
1809
+ 0,
1810
+ 1
1811
+ ]
1812
+ ],
1813
+ "flatten_loops": [
1814
+ true
1815
+ ],
1816
+ "l2_groupings": [
1817
+ 1
1818
+ ],
1819
+ "range_unroll_factors": [
1820
+ 0
1821
+ ],
1822
+ "range_warp_specializes": [],
1823
+ "range_num_stages": [
1824
+ 0
1825
+ ],
1826
+ "range_multi_buffers": [
1827
+ null
1828
+ ],
1829
+ "range_flattens": [
1830
+ null
1831
+ ],
1832
+ "load_eviction_policies": [
1833
+ "",
1834
+ "",
1835
+ ""
1836
+ ],
1837
+ "num_warps": 4,
1838
+ "num_stages": 1,
1839
+ "indexing": [
1840
+ "pointer",
1841
+ "pointer",
1842
+ "pointer",
1843
+ "pointer"
1844
+ ],
1845
+ "pid_type": "flat"
1846
+ },
1847
+ "intermediate_8192_numtokens_24": {
1848
+ "block_sizes": [
1849
+ 1,
1850
+ 2048
1851
+ ],
1852
+ "loop_orders": [
1853
+ [
1854
+ 0,
1855
+ 1
1856
+ ]
1857
+ ],
1858
+ "flatten_loops": [
1859
+ true
1860
+ ],
1861
+ "l2_groupings": [
1862
+ 32
1863
+ ],
1864
+ "range_unroll_factors": [
1865
+ 0
1866
+ ],
1867
+ "range_warp_specializes": [],
1868
+ "range_num_stages": [
1869
+ 0
1870
+ ],
1871
+ "range_multi_buffers": [
1872
+ null
1873
+ ],
1874
+ "range_flattens": [
1875
+ null
1876
+ ],
1877
+ "load_eviction_policies": [
1878
+ "",
1879
+ "last",
1880
+ "last"
1881
+ ],
1882
+ "num_warps": 32,
1883
+ "num_stages": 5,
1884
+ "indexing": [
1885
+ "pointer",
1886
+ "pointer",
1887
+ "pointer",
1888
+ "tensor_descriptor"
1889
+ ],
1890
+ "pid_type": "flat"
1891
+ },
1892
+ "intermediate_11008_numtokens_24": {
1893
+ "block_sizes": [
1894
+ 1,
1895
+ 1024
1896
+ ],
1897
+ "loop_orders": [
1898
+ [
1899
+ 1,
1900
+ 0
1901
+ ]
1902
+ ],
1903
+ "flatten_loops": [
1904
+ false
1905
+ ],
1906
+ "l2_groupings": [
1907
+ 64
1908
+ ],
1909
+ "range_unroll_factors": [
1910
+ 0
1911
+ ],
1912
+ "range_warp_specializes": [],
1913
+ "range_num_stages": [
1914
+ 0
1915
+ ],
1916
+ "range_multi_buffers": [
1917
+ null
1918
+ ],
1919
+ "range_flattens": [
1920
+ null
1921
+ ],
1922
+ "load_eviction_policies": [
1923
+ "last",
1924
+ "",
1925
+ ""
1926
+ ],
1927
+ "num_warps": 2,
1928
+ "num_stages": 2,
1929
+ "indexing": [
1930
+ "pointer",
1931
+ "pointer",
1932
+ "pointer",
1933
+ "pointer"
1934
+ ],
1935
+ "pid_type": "flat"
1936
+ },
1937
+ "intermediate_14336_numtokens_24": {
1938
+ "block_sizes": [
1939
+ 8,
1940
+ 512
1941
+ ],
1942
+ "loop_orders": [
1943
+ [
1944
+ 0,
1945
+ 1
1946
+ ]
1947
+ ],
1948
+ "flatten_loops": [
1949
+ true
1950
+ ],
1951
+ "l2_groupings": [
1952
+ 4
1953
+ ],
1954
+ "range_unroll_factors": [
1955
+ 0
1956
+ ],
1957
+ "range_warp_specializes": [],
1958
+ "range_num_stages": [
1959
+ 0
1960
+ ],
1961
+ "range_multi_buffers": [
1962
+ null
1963
+ ],
1964
+ "range_flattens": [
1965
+ null
1966
+ ],
1967
+ "load_eviction_policies": [
1968
+ "last",
1969
+ "first",
1970
+ "first"
1971
+ ],
1972
+ "num_warps": 32,
1973
+ "num_stages": 3,
1974
+ "indexing": [
1975
+ "pointer",
1976
+ "tensor_descriptor",
1977
+ "pointer",
1978
+ "tensor_descriptor"
1979
+ ],
1980
+ "pid_type": "flat"
1981
+ },
1982
+ "intermediate_2048_numtokens_32": {
1983
+ "block_sizes": [
1984
+ 32,
1985
+ 16
1986
+ ],
1987
+ "loop_orders": [
1988
+ [
1989
+ 0,
1990
+ 1
1991
+ ]
1992
+ ],
1993
+ "flatten_loops": [
1994
+ true
1995
+ ],
1996
+ "l2_groupings": [
1997
+ 1
1998
+ ],
1999
+ "range_unroll_factors": [
2000
+ 0
2001
+ ],
2002
+ "range_warp_specializes": [],
2003
+ "range_num_stages": [
2004
+ 0
2005
+ ],
2006
+ "range_multi_buffers": [
2007
+ null
2008
+ ],
2009
+ "range_flattens": [
2010
+ null
2011
+ ],
2012
+ "load_eviction_policies": [
2013
+ "",
2014
+ "",
2015
+ ""
2016
+ ],
2017
+ "num_warps": 4,
2018
+ "num_stages": 1,
2019
+ "indexing": [
2020
+ "pointer",
2021
+ "pointer",
2022
+ "pointer",
2023
+ "pointer"
2024
+ ],
2025
+ "pid_type": "flat"
2026
+ },
2027
+ "intermediate_2880_numtokens_32": {
2028
+ "block_sizes": [
2029
+ 4,
2030
+ 256
2031
+ ],
2032
+ "loop_orders": [
2033
+ [
2034
+ 0,
2035
+ 1
2036
+ ]
2037
+ ],
2038
+ "flatten_loops": [
2039
+ true
2040
+ ],
2041
+ "l2_groupings": [
2042
+ 1
2043
+ ],
2044
+ "range_unroll_factors": [
2045
+ 0
2046
+ ],
2047
+ "range_warp_specializes": [],
2048
+ "range_num_stages": [
2049
+ 0
2050
+ ],
2051
+ "range_multi_buffers": [
2052
+ null
2053
+ ],
2054
+ "range_flattens": [
2055
+ null
2056
+ ],
2057
+ "load_eviction_policies": [
2058
+ "",
2059
+ "",
2060
+ ""
2061
+ ],
2062
+ "num_warps": 4,
2063
+ "num_stages": 1,
2064
+ "indexing": [
2065
+ "pointer",
2066
+ "pointer",
2067
+ "pointer",
2068
+ "pointer"
2069
+ ],
2070
+ "pid_type": "flat"
2071
+ },
2072
+ "intermediate_4096_numtokens_32": {
2073
+ "block_sizes": [
2074
+ 4,
2075
+ 4096
2076
+ ],
2077
+ "loop_orders": [
2078
+ [
2079
+ 0,
2080
+ 1
2081
+ ]
2082
+ ],
2083
+ "flatten_loops": [
2084
+ true
2085
+ ],
2086
+ "l2_groupings": [
2087
+ 16
2088
+ ],
2089
+ "range_unroll_factors": [
2090
+ 0
2091
+ ],
2092
+ "range_warp_specializes": [],
2093
+ "range_num_stages": [
2094
+ 0
2095
+ ],
2096
+ "range_multi_buffers": [
2097
+ null
2098
+ ],
2099
+ "range_flattens": [
2100
+ null
2101
+ ],
2102
+ "load_eviction_policies": [
2103
+ "first",
2104
+ "last",
2105
+ ""
2106
+ ],
2107
+ "num_warps": 32,
2108
+ "num_stages": 5,
2109
+ "indexing": [
2110
+ "pointer",
2111
+ "tensor_descriptor",
2112
+ "pointer",
2113
+ "pointer"
2114
+ ],
2115
+ "pid_type": "flat"
2116
+ },
2117
+ "intermediate_8192_numtokens_32": {
2118
+ "block_sizes": [
2119
+ 4,
2120
+ 1024
2121
+ ],
2122
+ "loop_orders": [
2123
+ [
2124
+ 0,
2125
+ 1
2126
+ ]
2127
+ ],
2128
+ "flatten_loops": [
2129
+ true
2130
+ ],
2131
+ "l2_groupings": [
2132
+ 2
2133
+ ],
2134
+ "range_unroll_factors": [
2135
+ 0
2136
+ ],
2137
+ "range_warp_specializes": [],
2138
+ "range_num_stages": [
2139
+ 0
2140
+ ],
2141
+ "range_multi_buffers": [
2142
+ null
2143
+ ],
2144
+ "range_flattens": [
2145
+ null
2146
+ ],
2147
+ "load_eviction_policies": [
2148
+ "",
2149
+ "last",
2150
+ ""
2151
+ ],
2152
+ "num_warps": 2,
2153
+ "num_stages": 3,
2154
+ "indexing": [
2155
+ "tensor_descriptor",
2156
+ "tensor_descriptor",
2157
+ "pointer",
2158
+ "tensor_descriptor"
2159
+ ],
2160
+ "pid_type": "flat"
2161
+ },
2162
+ "intermediate_11008_numtokens_32": {
2163
+ "block_sizes": [
2164
+ 2,
2165
+ 2048
2166
+ ],
2167
+ "loop_orders": [
2168
+ [
2169
+ 0,
2170
+ 1
2171
+ ]
2172
+ ],
2173
+ "flatten_loops": [
2174
+ true
2175
+ ],
2176
+ "l2_groupings": [
2177
+ 1
2178
+ ],
2179
+ "range_unroll_factors": [
2180
+ 0
2181
+ ],
2182
+ "range_warp_specializes": [],
2183
+ "range_num_stages": [
2184
+ 0
2185
+ ],
2186
+ "range_multi_buffers": [
2187
+ null
2188
+ ],
2189
+ "range_flattens": [
2190
+ null
2191
+ ],
2192
+ "load_eviction_policies": [
2193
+ "",
2194
+ "",
2195
+ ""
2196
+ ],
2197
+ "num_warps": 8,
2198
+ "num_stages": 1,
2199
+ "indexing": [
2200
+ "pointer",
2201
+ "pointer",
2202
+ "pointer",
2203
+ "pointer"
2204
+ ],
2205
+ "pid_type": "flat"
2206
+ },
2207
+ "intermediate_14336_numtokens_32": {
2208
+ "block_sizes": [
2209
+ 1,
2210
+ 512
2211
+ ],
2212
+ "loop_orders": [
2213
+ [
2214
+ 0,
2215
+ 1
2216
+ ]
2217
+ ],
2218
+ "flatten_loops": [
2219
+ true
2220
+ ],
2221
+ "l2_groupings": [
2222
+ 64
2223
+ ],
2224
+ "range_unroll_factors": [
2225
+ 0
2226
+ ],
2227
+ "range_warp_specializes": [],
2228
+ "range_num_stages": [
2229
+ 0
2230
+ ],
2231
+ "range_multi_buffers": [
2232
+ null
2233
+ ],
2234
+ "range_flattens": [
2235
+ null
2236
+ ],
2237
+ "load_eviction_policies": [
2238
+ "first",
2239
+ "",
2240
+ ""
2241
+ ],
2242
+ "num_warps": 4,
2243
+ "num_stages": 3,
2244
+ "indexing": [
2245
+ "pointer",
2246
+ "tensor_descriptor",
2247
+ "pointer",
2248
+ "pointer"
2249
+ ],
2250
+ "pid_type": "flat"
2251
+ },
2252
+ "intermediate_2048_numtokens_40": {
2253
+ "block_sizes": [
2254
+ 32,
2255
+ 32
2256
+ ],
2257
+ "loop_orders": [
2258
+ [
2259
+ 0,
2260
+ 1
2261
+ ]
2262
+ ],
2263
+ "flatten_loops": [
2264
+ true
2265
+ ],
2266
+ "l2_groupings": [
2267
+ 1
2268
+ ],
2269
+ "range_unroll_factors": [
2270
+ 0
2271
+ ],
2272
+ "range_warp_specializes": [],
2273
+ "range_num_stages": [
2274
+ 0
2275
+ ],
2276
+ "range_multi_buffers": [
2277
+ null
2278
+ ],
2279
+ "range_flattens": [
2280
+ null
2281
+ ],
2282
+ "load_eviction_policies": [
2283
+ "",
2284
+ "",
2285
+ ""
2286
+ ],
2287
+ "num_warps": 4,
2288
+ "num_stages": 1,
2289
+ "indexing": [
2290
+ "pointer",
2291
+ "pointer",
2292
+ "pointer",
2293
+ "pointer"
2294
+ ],
2295
+ "pid_type": "flat"
2296
+ },
2297
+ "intermediate_2880_numtokens_40": {
2298
+ "block_sizes": [
2299
+ 1,
2300
+ 4096
2301
+ ],
2302
+ "loop_orders": [
2303
+ [
2304
+ 1,
2305
+ 0
2306
+ ]
2307
+ ],
2308
+ "flatten_loops": [
2309
+ false
2310
+ ],
2311
+ "l2_groupings": [
2312
+ 32
2313
+ ],
2314
+ "range_unroll_factors": [
2315
+ 0
2316
+ ],
2317
+ "range_warp_specializes": [],
2318
+ "range_num_stages": [
2319
+ 0
2320
+ ],
2321
+ "range_multi_buffers": [
2322
+ null
2323
+ ],
2324
+ "range_flattens": [
2325
+ null
2326
+ ],
2327
+ "load_eviction_policies": [
2328
+ "",
2329
+ "last",
2330
+ "last"
2331
+ ],
2332
+ "num_warps": 8,
2333
+ "num_stages": 4,
2334
+ "indexing": [
2335
+ "tensor_descriptor",
2336
+ "pointer",
2337
+ "pointer",
2338
+ "pointer"
2339
+ ],
2340
+ "pid_type": "flat"
2341
+ },
2342
+ "intermediate_4096_numtokens_40": {
2343
+ "block_sizes": [
2344
+ 32,
2345
+ 64
2346
+ ],
2347
+ "loop_orders": [
2348
+ [
2349
+ 0,
2350
+ 1
2351
+ ]
2352
+ ],
2353
+ "flatten_loops": [
2354
+ true
2355
+ ],
2356
+ "l2_groupings": [
2357
+ 1
2358
+ ],
2359
+ "range_unroll_factors": [
2360
+ 0
2361
+ ],
2362
+ "range_warp_specializes": [],
2363
+ "range_num_stages": [
2364
+ 0
2365
+ ],
2366
+ "range_multi_buffers": [
2367
+ null
2368
+ ],
2369
+ "range_flattens": [
2370
+ null
2371
+ ],
2372
+ "load_eviction_policies": [
2373
+ "",
2374
+ "",
2375
+ ""
2376
+ ],
2377
+ "num_warps": 8,
2378
+ "num_stages": 1,
2379
+ "indexing": [
2380
+ "pointer",
2381
+ "pointer",
2382
+ "pointer",
2383
+ "pointer"
2384
+ ],
2385
+ "pid_type": "flat"
2386
+ },
2387
+ "intermediate_8192_numtokens_40": {
2388
+ "block_sizes": [
2389
+ 2,
2390
+ 256
2391
+ ],
2392
+ "loop_orders": [
2393
+ [
2394
+ 0,
2395
+ 1
2396
+ ]
2397
+ ],
2398
+ "flatten_loops": [
2399
+ true
2400
+ ],
2401
+ "l2_groupings": [
2402
+ 16
2403
+ ],
2404
+ "range_unroll_factors": [
2405
+ 0
2406
+ ],
2407
+ "range_warp_specializes": [],
2408
+ "range_num_stages": [
2409
+ 0
2410
+ ],
2411
+ "range_multi_buffers": [
2412
+ null
2413
+ ],
2414
+ "range_flattens": [
2415
+ null
2416
+ ],
2417
+ "load_eviction_policies": [
2418
+ "",
2419
+ "last",
2420
+ "last"
2421
+ ],
2422
+ "num_warps": 2,
2423
+ "num_stages": 2,
2424
+ "indexing": [
2425
+ "pointer",
2426
+ "pointer",
2427
+ "pointer",
2428
+ "tensor_descriptor"
2429
+ ],
2430
+ "pid_type": "flat"
2431
+ },
2432
+ "intermediate_11008_numtokens_40": {
2433
+ "block_sizes": [
2434
+ 16,
2435
+ 256
2436
+ ],
2437
+ "loop_orders": [
2438
+ [
2439
+ 0,
2440
+ 1
2441
+ ]
2442
+ ],
2443
+ "flatten_loops": [
2444
+ true
2445
+ ],
2446
+ "l2_groupings": [
2447
+ 1
2448
+ ],
2449
+ "range_unroll_factors": [
2450
+ 0
2451
+ ],
2452
+ "range_warp_specializes": [],
2453
+ "range_num_stages": [
2454
+ 0
2455
+ ],
2456
+ "range_multi_buffers": [
2457
+ null
2458
+ ],
2459
+ "range_flattens": [
2460
+ null
2461
+ ],
2462
+ "load_eviction_policies": [
2463
+ "",
2464
+ "",
2465
+ ""
2466
+ ],
2467
+ "num_warps": 8,
2468
+ "num_stages": 1,
2469
+ "indexing": [
2470
+ "pointer",
2471
+ "pointer",
2472
+ "pointer",
2473
+ "pointer"
2474
+ ],
2475
+ "pid_type": "flat"
2476
+ },
2477
+ "intermediate_14336_numtokens_40": {
2478
+ "block_sizes": [
2479
+ 4,
2480
+ 1024
2481
+ ],
2482
+ "loop_orders": [
2483
+ [
2484
+ 0,
2485
+ 1
2486
+ ]
2487
+ ],
2488
+ "flatten_loops": [
2489
+ true
2490
+ ],
2491
+ "l2_groupings": [
2492
+ 1
2493
+ ],
2494
+ "range_unroll_factors": [
2495
+ 0
2496
+ ],
2497
+ "range_warp_specializes": [],
2498
+ "range_num_stages": [
2499
+ 1
2500
+ ],
2501
+ "range_multi_buffers": [
2502
+ null
2503
+ ],
2504
+ "range_flattens": [
2505
+ null
2506
+ ],
2507
+ "load_eviction_policies": [
2508
+ "first",
2509
+ "",
2510
+ "last"
2511
+ ],
2512
+ "num_warps": 16,
2513
+ "num_stages": 5,
2514
+ "indexing": [
2515
+ "tensor_descriptor",
2516
+ "tensor_descriptor",
2517
+ "tensor_descriptor",
2518
+ "tensor_descriptor"
2519
+ ],
2520
+ "pid_type": "persistent_interleaved",
2521
+ "num_sm_multiplier": 32,
2522
+ "maxnreg": 32
2523
+ },
2524
+ "intermediate_2048_numtokens_48": {
2525
+ "block_sizes": [
2526
+ 32,
2527
+ 32
2528
+ ],
2529
+ "loop_orders": [
2530
+ [
2531
+ 0,
2532
+ 1
2533
+ ]
2534
+ ],
2535
+ "flatten_loops": [
2536
+ true
2537
+ ],
2538
+ "l2_groupings": [
2539
+ 1
2540
+ ],
2541
+ "range_unroll_factors": [
2542
+ 0
2543
+ ],
2544
+ "range_warp_specializes": [],
2545
+ "range_num_stages": [
2546
+ 0
2547
+ ],
2548
+ "range_multi_buffers": [
2549
+ null
2550
+ ],
2551
+ "range_flattens": [
2552
+ null
2553
+ ],
2554
+ "load_eviction_policies": [
2555
+ "",
2556
+ "",
2557
+ ""
2558
+ ],
2559
+ "num_warps": 4,
2560
+ "num_stages": 1,
2561
+ "indexing": [
2562
+ "pointer",
2563
+ "pointer",
2564
+ "pointer",
2565
+ "pointer"
2566
+ ],
2567
+ "pid_type": "flat"
2568
+ },
2569
+ "intermediate_2880_numtokens_48": {
2570
+ "block_sizes": [
2571
+ 16,
2572
+ 64
2573
+ ],
2574
+ "loop_orders": [
2575
+ [
2576
+ 0,
2577
+ 1
2578
+ ]
2579
+ ],
2580
+ "flatten_loops": [
2581
+ true
2582
+ ],
2583
+ "l2_groupings": [
2584
+ 1
2585
+ ],
2586
+ "range_unroll_factors": [
2587
+ 0
2588
+ ],
2589
+ "range_warp_specializes": [],
2590
+ "range_num_stages": [
2591
+ 0
2592
+ ],
2593
+ "range_multi_buffers": [
2594
+ null
2595
+ ],
2596
+ "range_flattens": [
2597
+ null
2598
+ ],
2599
+ "load_eviction_policies": [
2600
+ "",
2601
+ "",
2602
+ ""
2603
+ ],
2604
+ "num_warps": 4,
2605
+ "num_stages": 1,
2606
+ "indexing": [
2607
+ "pointer",
2608
+ "pointer",
2609
+ "pointer",
2610
+ "pointer"
2611
+ ],
2612
+ "pid_type": "flat"
2613
+ },
2614
+ "intermediate_4096_numtokens_48": {
2615
+ "block_sizes": [
2616
+ 32,
2617
+ 64
2618
+ ],
2619
+ "loop_orders": [
2620
+ [
2621
+ 0,
2622
+ 1
2623
+ ]
2624
+ ],
2625
+ "flatten_loops": [
2626
+ true
2627
+ ],
2628
+ "l2_groupings": [
2629
+ 1
2630
+ ],
2631
+ "range_unroll_factors": [
2632
+ 0
2633
+ ],
2634
+ "range_warp_specializes": [],
2635
+ "range_num_stages": [
2636
+ 0
2637
+ ],
2638
+ "range_multi_buffers": [
2639
+ null
2640
+ ],
2641
+ "range_flattens": [
2642
+ null
2643
+ ],
2644
+ "load_eviction_policies": [
2645
+ "",
2646
+ "",
2647
+ ""
2648
+ ],
2649
+ "num_warps": 8,
2650
+ "num_stages": 1,
2651
+ "indexing": [
2652
+ "pointer",
2653
+ "pointer",
2654
+ "pointer",
2655
+ "pointer"
2656
+ ],
2657
+ "pid_type": "flat"
2658
+ },
2659
+ "intermediate_8192_numtokens_48": {
2660
+ "block_sizes": [
2661
+ 1,
2662
+ 4096
2663
+ ],
2664
+ "loop_orders": [
2665
+ [
2666
+ 0,
2667
+ 1
2668
+ ]
2669
+ ],
2670
+ "flatten_loops": [
2671
+ true
2672
+ ],
2673
+ "l2_groupings": [
2674
+ 2
2675
+ ],
2676
+ "range_unroll_factors": [
2677
+ 0
2678
+ ],
2679
+ "range_warp_specializes": [],
2680
+ "range_num_stages": [
2681
+ 0
2682
+ ],
2683
+ "range_multi_buffers": [
2684
+ null
2685
+ ],
2686
+ "range_flattens": [
2687
+ null
2688
+ ],
2689
+ "load_eviction_policies": [
2690
+ "first",
2691
+ "",
2692
+ "last"
2693
+ ],
2694
+ "num_warps": 4,
2695
+ "num_stages": 2,
2696
+ "indexing": [
2697
+ "pointer",
2698
+ "tensor_descriptor",
2699
+ "tensor_descriptor",
2700
+ "pointer"
2701
+ ],
2702
+ "pid_type": "flat"
2703
+ },
2704
+ "intermediate_11008_numtokens_48": {
2705
+ "block_sizes": [
2706
+ 16,
2707
+ 256
2708
+ ],
2709
+ "loop_orders": [
2710
+ [
2711
+ 0,
2712
+ 1
2713
+ ]
2714
+ ],
2715
+ "flatten_loops": [
2716
+ true
2717
+ ],
2718
+ "l2_groupings": [
2719
+ 1
2720
+ ],
2721
+ "range_unroll_factors": [
2722
+ 0
2723
+ ],
2724
+ "range_warp_specializes": [],
2725
+ "range_num_stages": [
2726
+ 0
2727
+ ],
2728
+ "range_multi_buffers": [
2729
+ null
2730
+ ],
2731
+ "range_flattens": [
2732
+ null
2733
+ ],
2734
+ "load_eviction_policies": [
2735
+ "",
2736
+ "",
2737
+ ""
2738
+ ],
2739
+ "num_warps": 8,
2740
+ "num_stages": 1,
2741
+ "indexing": [
2742
+ "pointer",
2743
+ "pointer",
2744
+ "pointer",
2745
+ "pointer"
2746
+ ],
2747
+ "pid_type": "flat"
2748
+ },
2749
+ "intermediate_14336_numtokens_48": {
2750
+ "block_sizes": [
2751
+ 32,
2752
+ 256
2753
+ ],
2754
+ "loop_orders": [
2755
+ [
2756
+ 0,
2757
+ 1
2758
+ ]
2759
+ ],
2760
+ "flatten_loops": [
2761
+ true
2762
+ ],
2763
+ "l2_groupings": [
2764
+ 1
2765
+ ],
2766
+ "range_unroll_factors": [
2767
+ 0
2768
+ ],
2769
+ "range_warp_specializes": [],
2770
+ "range_num_stages": [
2771
+ 0
2772
+ ],
2773
+ "range_multi_buffers": [
2774
+ null
2775
+ ],
2776
+ "range_flattens": [
2777
+ null
2778
+ ],
2779
+ "load_eviction_policies": [
2780
+ "",
2781
+ "",
2782
+ ""
2783
+ ],
2784
+ "num_warps": 8,
2785
+ "num_stages": 1,
2786
+ "indexing": [
2787
+ "pointer",
2788
+ "pointer",
2789
+ "pointer",
2790
+ "pointer"
2791
+ ],
2792
+ "pid_type": "flat"
2793
+ },
2794
+ "intermediate_2048_numtokens_56": {
2795
+ "block_sizes": [
2796
+ 32,
2797
+ 32
2798
+ ],
2799
+ "loop_orders": [
2800
+ [
2801
+ 0,
2802
+ 1
2803
+ ]
2804
+ ],
2805
+ "flatten_loops": [
2806
+ true
2807
+ ],
2808
+ "l2_groupings": [
2809
+ 1
2810
+ ],
2811
+ "range_unroll_factors": [
2812
+ 0
2813
+ ],
2814
+ "range_warp_specializes": [],
2815
+ "range_num_stages": [
2816
+ 0
2817
+ ],
2818
+ "range_multi_buffers": [
2819
+ null
2820
+ ],
2821
+ "range_flattens": [
2822
+ null
2823
+ ],
2824
+ "load_eviction_policies": [
2825
+ "",
2826
+ "",
2827
+ ""
2828
+ ],
2829
+ "num_warps": 4,
2830
+ "num_stages": 1,
2831
+ "indexing": [
2832
+ "pointer",
2833
+ "pointer",
2834
+ "pointer",
2835
+ "pointer"
2836
+ ],
2837
+ "pid_type": "flat"
2838
+ },
2839
+ "intermediate_2880_numtokens_56": {
2840
+ "block_sizes": [
2841
+ 1,
2842
+ 2048
2843
+ ],
2844
+ "loop_orders": [
2845
+ [
2846
+ 0,
2847
+ 1
2848
+ ]
2849
+ ],
2850
+ "flatten_loops": [
2851
+ true
2852
+ ],
2853
+ "l2_groupings": [
2854
+ 1
2855
+ ],
2856
+ "range_unroll_factors": [
2857
+ 0
2858
+ ],
2859
+ "range_warp_specializes": [],
2860
+ "range_num_stages": [
2861
+ 0
2862
+ ],
2863
+ "range_multi_buffers": [
2864
+ null
2865
+ ],
2866
+ "range_flattens": [
2867
+ null
2868
+ ],
2869
+ "load_eviction_policies": [
2870
+ "",
2871
+ "",
2872
+ ""
2873
+ ],
2874
+ "num_warps": 8,
2875
+ "num_stages": 1,
2876
+ "indexing": [
2877
+ "pointer",
2878
+ "pointer",
2879
+ "pointer",
2880
+ "pointer"
2881
+ ],
2882
+ "pid_type": "flat"
2883
+ },
2884
+ "intermediate_4096_numtokens_56": {
2885
+ "block_sizes": [
2886
+ 32,
2887
+ 64
2888
+ ],
2889
+ "loop_orders": [
2890
+ [
2891
+ 0,
2892
+ 1
2893
+ ]
2894
+ ],
2895
+ "flatten_loops": [
2896
+ true
2897
+ ],
2898
+ "l2_groupings": [
2899
+ 1
2900
+ ],
2901
+ "range_unroll_factors": [
2902
+ 0
2903
+ ],
2904
+ "range_warp_specializes": [],
2905
+ "range_num_stages": [
2906
+ 0
2907
+ ],
2908
+ "range_multi_buffers": [
2909
+ null
2910
+ ],
2911
+ "range_flattens": [
2912
+ null
2913
+ ],
2914
+ "load_eviction_policies": [
2915
+ "",
2916
+ "",
2917
+ ""
2918
+ ],
2919
+ "num_warps": 8,
2920
+ "num_stages": 1,
2921
+ "indexing": [
2922
+ "pointer",
2923
+ "pointer",
2924
+ "pointer",
2925
+ "pointer"
2926
+ ],
2927
+ "pid_type": "flat"
2928
+ },
2929
+ "intermediate_8192_numtokens_56": {
2930
+ "block_sizes": [
2931
+ 32,
2932
+ 128
2933
+ ],
2934
+ "loop_orders": [
2935
+ [
2936
+ 0,
2937
+ 1
2938
+ ]
2939
+ ],
2940
+ "flatten_loops": [
2941
+ true
2942
+ ],
2943
+ "l2_groupings": [
2944
+ 1
2945
+ ],
2946
+ "range_unroll_factors": [
2947
+ 0
2948
+ ],
2949
+ "range_warp_specializes": [],
2950
+ "range_num_stages": [
2951
+ 0
2952
+ ],
2953
+ "range_multi_buffers": [
2954
+ null
2955
+ ],
2956
+ "range_flattens": [
2957
+ null
2958
+ ],
2959
+ "load_eviction_policies": [
2960
+ "",
2961
+ "",
2962
+ ""
2963
+ ],
2964
+ "num_warps": 8,
2965
+ "num_stages": 1,
2966
+ "indexing": [
2967
+ "pointer",
2968
+ "pointer",
2969
+ "pointer",
2970
+ "pointer"
2971
+ ],
2972
+ "pid_type": "flat"
2973
+ },
2974
+ "intermediate_11008_numtokens_56": {
2975
+ "block_sizes": [
2976
+ 1,
2977
+ 8192
2978
+ ],
2979
+ "loop_orders": [
2980
+ [
2981
+ 0,
2982
+ 1
2983
+ ]
2984
+ ],
2985
+ "flatten_loops": [
2986
+ true
2987
+ ],
2988
+ "l2_groupings": [
2989
+ 1
2990
+ ],
2991
+ "range_unroll_factors": [
2992
+ 0
2993
+ ],
2994
+ "range_warp_specializes": [],
2995
+ "range_num_stages": [
2996
+ 0
2997
+ ],
2998
+ "range_multi_buffers": [
2999
+ null
3000
+ ],
3001
+ "range_flattens": [
3002
+ null
3003
+ ],
3004
+ "load_eviction_policies": [
3005
+ "",
3006
+ "",
3007
+ ""
3008
+ ],
3009
+ "num_warps": 8,
3010
+ "num_stages": 1,
3011
+ "indexing": [
3012
+ "pointer",
3013
+ "pointer",
3014
+ "pointer",
3015
+ "pointer"
3016
+ ],
3017
+ "pid_type": "flat"
3018
+ },
3019
+ "intermediate_14336_numtokens_56": {
3020
+ "block_sizes": [
3021
+ 2,
3022
+ 4096
3023
+ ],
3024
+ "loop_orders": [
3025
+ [
3026
+ 0,
3027
+ 1
3028
+ ]
3029
+ ],
3030
+ "flatten_loops": [
3031
+ false
3032
+ ],
3033
+ "l2_groupings": [
3034
+ 32
3035
+ ],
3036
+ "range_unroll_factors": [
3037
+ 0
3038
+ ],
3039
+ "range_warp_specializes": [],
3040
+ "range_num_stages": [
3041
+ 0
3042
+ ],
3043
+ "range_multi_buffers": [
3044
+ null
3045
+ ],
3046
+ "range_flattens": [
3047
+ null
3048
+ ],
3049
+ "load_eviction_policies": [
3050
+ "first",
3051
+ "first",
3052
+ ""
3053
+ ],
3054
+ "num_warps": 2,
3055
+ "num_stages": 4,
3056
+ "indexing": [
3057
+ "tensor_descriptor",
3058
+ "pointer",
3059
+ "tensor_descriptor",
3060
+ "tensor_descriptor"
3061
+ ],
3062
+ "pid_type": "flat"
3063
+ },
3064
+ "intermediate_2048_numtokens_64": {
3065
+ "block_sizes": [
3066
+ 64,
3067
+ 32
3068
+ ],
3069
+ "loop_orders": [
3070
+ [
3071
+ 0,
3072
+ 1
3073
+ ]
3074
+ ],
3075
+ "flatten_loops": [
3076
+ true
3077
+ ],
3078
+ "l2_groupings": [
3079
+ 1
3080
+ ],
3081
+ "range_unroll_factors": [
3082
+ 0
3083
+ ],
3084
+ "range_warp_specializes": [],
3085
+ "range_num_stages": [
3086
+ 0
3087
+ ],
3088
+ "range_multi_buffers": [
3089
+ null
3090
+ ],
3091
+ "range_flattens": [
3092
+ null
3093
+ ],
3094
+ "load_eviction_policies": [
3095
+ "",
3096
+ "",
3097
+ ""
3098
+ ],
3099
+ "num_warps": 8,
3100
+ "num_stages": 1,
3101
+ "indexing": [
3102
+ "pointer",
3103
+ "pointer",
3104
+ "pointer",
3105
+ "pointer"
3106
+ ],
3107
+ "pid_type": "flat"
3108
+ },
3109
+ "intermediate_2880_numtokens_64": {
3110
+ "block_sizes": [
3111
+ 1,
3112
+ 2048
3113
+ ],
3114
+ "loop_orders": [
3115
+ [
3116
+ 0,
3117
+ 1
3118
+ ]
3119
+ ],
3120
+ "flatten_loops": [
3121
+ true
3122
+ ],
3123
+ "l2_groupings": [
3124
+ 1
3125
+ ],
3126
+ "range_unroll_factors": [
3127
+ 0
3128
+ ],
3129
+ "range_warp_specializes": [],
3130
+ "range_num_stages": [
3131
+ 0
3132
+ ],
3133
+ "range_multi_buffers": [
3134
+ null
3135
+ ],
3136
+ "range_flattens": [
3137
+ null
3138
+ ],
3139
+ "load_eviction_policies": [
3140
+ "",
3141
+ "",
3142
+ ""
3143
+ ],
3144
+ "num_warps": 8,
3145
+ "num_stages": 1,
3146
+ "indexing": [
3147
+ "pointer",
3148
+ "pointer",
3149
+ "pointer",
3150
+ "pointer"
3151
+ ],
3152
+ "pid_type": "flat"
3153
+ },
3154
+ "intermediate_4096_numtokens_64": {
3155
+ "block_sizes": [
3156
+ 64,
3157
+ 32
3158
+ ],
3159
+ "loop_orders": [
3160
+ [
3161
+ 0,
3162
+ 1
3163
+ ]
3164
+ ],
3165
+ "flatten_loops": [
3166
+ true
3167
+ ],
3168
+ "l2_groupings": [
3169
+ 1
3170
+ ],
3171
+ "range_unroll_factors": [
3172
+ 0
3173
+ ],
3174
+ "range_warp_specializes": [],
3175
+ "range_num_stages": [
3176
+ 0
3177
+ ],
3178
+ "range_multi_buffers": [
3179
+ null
3180
+ ],
3181
+ "range_flattens": [
3182
+ null
3183
+ ],
3184
+ "load_eviction_policies": [
3185
+ "",
3186
+ "",
3187
+ ""
3188
+ ],
3189
+ "num_warps": 8,
3190
+ "num_stages": 1,
3191
+ "indexing": [
3192
+ "pointer",
3193
+ "pointer",
3194
+ "pointer",
3195
+ "pointer"
3196
+ ],
3197
+ "pid_type": "flat"
3198
+ },
3199
+ "intermediate_8192_numtokens_64": {
3200
+ "block_sizes": [
3201
+ 64,
3202
+ 64
3203
+ ],
3204
+ "loop_orders": [
3205
+ [
3206
+ 0,
3207
+ 1
3208
+ ]
3209
+ ],
3210
+ "flatten_loops": [
3211
+ true
3212
+ ],
3213
+ "l2_groupings": [
3214
+ 1
3215
+ ],
3216
+ "range_unroll_factors": [
3217
+ 0
3218
+ ],
3219
+ "range_warp_specializes": [],
3220
+ "range_num_stages": [
3221
+ 0
3222
+ ],
3223
+ "range_multi_buffers": [
3224
+ null
3225
+ ],
3226
+ "range_flattens": [
3227
+ null
3228
+ ],
3229
+ "load_eviction_policies": [
3230
+ "",
3231
+ "",
3232
+ ""
3233
+ ],
3234
+ "num_warps": 8,
3235
+ "num_stages": 1,
3236
+ "indexing": [
3237
+ "pointer",
3238
+ "pointer",
3239
+ "pointer",
3240
+ "pointer"
3241
+ ],
3242
+ "pid_type": "flat"
3243
+ },
3244
+ "intermediate_11008_numtokens_64": {
3245
+ "block_sizes": [
3246
+ 1,
3247
+ 8192
3248
+ ],
3249
+ "loop_orders": [
3250
+ [
3251
+ 0,
3252
+ 1
3253
+ ]
3254
+ ],
3255
+ "flatten_loops": [
3256
+ true
3257
+ ],
3258
+ "l2_groupings": [
3259
+ 1
3260
+ ],
3261
+ "range_unroll_factors": [
3262
+ 0
3263
+ ],
3264
+ "range_warp_specializes": [],
3265
+ "range_num_stages": [
3266
+ 0
3267
+ ],
3268
+ "range_multi_buffers": [
3269
+ null
3270
+ ],
3271
+ "range_flattens": [
3272
+ null
3273
+ ],
3274
+ "load_eviction_policies": [
3275
+ "",
3276
+ "",
3277
+ ""
3278
+ ],
3279
+ "num_warps": 8,
3280
+ "num_stages": 1,
3281
+ "indexing": [
3282
+ "pointer",
3283
+ "pointer",
3284
+ "pointer",
3285
+ "pointer"
3286
+ ],
3287
+ "pid_type": "flat"
3288
+ },
3289
+ "intermediate_14336_numtokens_64": {
3290
+ "block_sizes": [
3291
+ 16,
3292
+ 512
3293
+ ],
3294
+ "loop_orders": [
3295
+ [
3296
+ 0,
3297
+ 1
3298
+ ]
3299
+ ],
3300
+ "flatten_loops": [
3301
+ true
3302
+ ],
3303
+ "l2_groupings": [
3304
+ 16
3305
+ ],
3306
+ "range_unroll_factors": [
3307
+ 0
3308
+ ],
3309
+ "range_warp_specializes": [],
3310
+ "range_num_stages": [
3311
+ 0
3312
+ ],
3313
+ "range_multi_buffers": [
3314
+ null
3315
+ ],
3316
+ "range_flattens": [
3317
+ null
3318
+ ],
3319
+ "load_eviction_policies": [
3320
+ "",
3321
+ "first",
3322
+ ""
3323
+ ],
3324
+ "num_warps": 4,
3325
+ "num_stages": 1,
3326
+ "indexing": [
3327
+ "pointer",
3328
+ "tensor_descriptor",
3329
+ "tensor_descriptor",
3330
+ "tensor_descriptor"
3331
+ ],
3332
+ "pid_type": "flat"
3333
+ },
3334
+ "intermediate_2048_numtokens_72": {
3335
+ "block_sizes": [
3336
+ 64,
3337
+ 32
3338
+ ],
3339
+ "loop_orders": [
3340
+ [
3341
+ 0,
3342
+ 1
3343
+ ]
3344
+ ],
3345
+ "flatten_loops": [
3346
+ true
3347
+ ],
3348
+ "l2_groupings": [
3349
+ 1
3350
+ ],
3351
+ "range_unroll_factors": [
3352
+ 0
3353
+ ],
3354
+ "range_warp_specializes": [],
3355
+ "range_num_stages": [
3356
+ 0
3357
+ ],
3358
+ "range_multi_buffers": [
3359
+ null
3360
+ ],
3361
+ "range_flattens": [
3362
+ null
3363
+ ],
3364
+ "load_eviction_policies": [
3365
+ "",
3366
+ "",
3367
+ ""
3368
+ ],
3369
+ "num_warps": 8,
3370
+ "num_stages": 1,
3371
+ "indexing": [
3372
+ "pointer",
3373
+ "pointer",
3374
+ "pointer",
3375
+ "pointer"
3376
+ ],
3377
+ "pid_type": "flat"
3378
+ },
3379
+ "intermediate_2880_numtokens_72": {
3380
+ "block_sizes": [
3381
+ 32,
3382
+ 64
3383
+ ],
3384
+ "loop_orders": [
3385
+ [
3386
+ 0,
3387
+ 1
3388
+ ]
3389
+ ],
3390
+ "flatten_loops": [
3391
+ true
3392
+ ],
3393
+ "l2_groupings": [
3394
+ 1
3395
+ ],
3396
+ "range_unroll_factors": [
3397
+ 0
3398
+ ],
3399
+ "range_warp_specializes": [],
3400
+ "range_num_stages": [
3401
+ 0
3402
+ ],
3403
+ "range_multi_buffers": [
3404
+ null
3405
+ ],
3406
+ "range_flattens": [
3407
+ null
3408
+ ],
3409
+ "load_eviction_policies": [
3410
+ "",
3411
+ "",
3412
+ ""
3413
+ ],
3414
+ "num_warps": 8,
3415
+ "num_stages": 1,
3416
+ "indexing": [
3417
+ "pointer",
3418
+ "pointer",
3419
+ "pointer",
3420
+ "pointer"
3421
+ ],
3422
+ "pid_type": "flat"
3423
+ },
3424
+ "intermediate_4096_numtokens_72": {
3425
+ "block_sizes": [
3426
+ 1,
3427
+ 1024
3428
+ ],
3429
+ "loop_orders": [
3430
+ [
3431
+ 0,
3432
+ 1
3433
+ ]
3434
+ ],
3435
+ "flatten_loops": [
3436
+ true
3437
+ ],
3438
+ "l2_groupings": [
3439
+ 4
3440
+ ],
3441
+ "range_unroll_factors": [
3442
+ 0
3443
+ ],
3444
+ "range_warp_specializes": [],
3445
+ "range_num_stages": [
3446
+ 0
3447
+ ],
3448
+ "range_multi_buffers": [
3449
+ null
3450
+ ],
3451
+ "range_flattens": [
3452
+ null
3453
+ ],
3454
+ "load_eviction_policies": [
3455
+ "",
3456
+ "",
3457
+ "first"
3458
+ ],
3459
+ "num_warps": 16,
3460
+ "num_stages": 2,
3461
+ "indexing": [
3462
+ "tensor_descriptor",
3463
+ "pointer",
3464
+ "pointer",
3465
+ "tensor_descriptor"
3466
+ ],
3467
+ "pid_type": "flat"
3468
+ },
3469
+ "intermediate_8192_numtokens_72": {
3470
+ "block_sizes": [
3471
+ 64,
3472
+ 128
3473
+ ],
3474
+ "loop_orders": [
3475
+ [
3476
+ 0,
3477
+ 1
3478
+ ]
3479
+ ],
3480
+ "flatten_loops": [
3481
+ true
3482
+ ],
3483
+ "l2_groupings": [
3484
+ 1
3485
+ ],
3486
+ "range_unroll_factors": [
3487
+ 0
3488
+ ],
3489
+ "range_warp_specializes": [],
3490
+ "range_num_stages": [
3491
+ 0
3492
+ ],
3493
+ "range_multi_buffers": [
3494
+ null
3495
+ ],
3496
+ "range_flattens": [
3497
+ null
3498
+ ],
3499
+ "load_eviction_policies": [
3500
+ "",
3501
+ "",
3502
+ ""
3503
+ ],
3504
+ "num_warps": 8,
3505
+ "num_stages": 1,
3506
+ "indexing": [
3507
+ "pointer",
3508
+ "pointer",
3509
+ "pointer",
3510
+ "pointer"
3511
+ ],
3512
+ "pid_type": "flat"
3513
+ },
3514
+ "intermediate_11008_numtokens_72": {
3515
+ "block_sizes": [
3516
+ 4,
3517
+ 256
3518
+ ],
3519
+ "loop_orders": [
3520
+ [
3521
+ 0,
3522
+ 1
3523
+ ]
3524
+ ],
3525
+ "flatten_loops": [
3526
+ true
3527
+ ],
3528
+ "l2_groupings": [
3529
+ 2
3530
+ ],
3531
+ "range_unroll_factors": [
3532
+ 0
3533
+ ],
3534
+ "range_warp_specializes": [],
3535
+ "range_num_stages": [
3536
+ 0
3537
+ ],
3538
+ "range_multi_buffers": [
3539
+ null
3540
+ ],
3541
+ "range_flattens": [
3542
+ null
3543
+ ],
3544
+ "load_eviction_policies": [
3545
+ "first",
3546
+ "last",
3547
+ "last"
3548
+ ],
3549
+ "num_warps": 32,
3550
+ "num_stages": 5,
3551
+ "indexing": [
3552
+ "pointer",
3553
+ "pointer",
3554
+ "tensor_descriptor",
3555
+ "tensor_descriptor"
3556
+ ],
3557
+ "pid_type": "flat"
3558
+ },
3559
+ "intermediate_14336_numtokens_72": {
3560
+ "block_sizes": [
3561
+ 32,
3562
+ 32
3563
+ ],
3564
+ "loop_orders": [
3565
+ [
3566
+ 0,
3567
+ 1
3568
+ ]
3569
+ ],
3570
+ "flatten_loops": [
3571
+ true
3572
+ ],
3573
+ "l2_groupings": [
3574
+ 2
3575
+ ],
3576
+ "range_unroll_factors": [
3577
+ 0
3578
+ ],
3579
+ "range_warp_specializes": [],
3580
+ "range_num_stages": [
3581
+ 0
3582
+ ],
3583
+ "range_multi_buffers": [
3584
+ null
3585
+ ],
3586
+ "range_flattens": [
3587
+ null
3588
+ ],
3589
+ "load_eviction_policies": [
3590
+ "last",
3591
+ "",
3592
+ "last"
3593
+ ],
3594
+ "num_warps": 8,
3595
+ "num_stages": 3,
3596
+ "indexing": [
3597
+ "pointer",
3598
+ "pointer",
3599
+ "tensor_descriptor",
3600
+ "pointer"
3601
+ ],
3602
+ "pid_type": "flat"
3603
+ },
3604
+ "intermediate_2048_numtokens_80": {
3605
+ "block_sizes": [
3606
+ 64,
3607
+ 32
3608
+ ],
3609
+ "loop_orders": [
3610
+ [
3611
+ 0,
3612
+ 1
3613
+ ]
3614
+ ],
3615
+ "flatten_loops": [
3616
+ true
3617
+ ],
3618
+ "l2_groupings": [
3619
+ 1
3620
+ ],
3621
+ "range_unroll_factors": [
3622
+ 0
3623
+ ],
3624
+ "range_warp_specializes": [],
3625
+ "range_num_stages": [
3626
+ 0
3627
+ ],
3628
+ "range_multi_buffers": [
3629
+ null
3630
+ ],
3631
+ "range_flattens": [
3632
+ null
3633
+ ],
3634
+ "load_eviction_policies": [
3635
+ "",
3636
+ "",
3637
+ ""
3638
+ ],
3639
+ "num_warps": 8,
3640
+ "num_stages": 1,
3641
+ "indexing": [
3642
+ "pointer",
3643
+ "pointer",
3644
+ "pointer",
3645
+ "pointer"
3646
+ ],
3647
+ "pid_type": "flat"
3648
+ },
3649
+ "intermediate_2880_numtokens_80": {
3650
+ "block_sizes": [
3651
+ 32,
3652
+ 64
3653
+ ],
3654
+ "loop_orders": [
3655
+ [
3656
+ 0,
3657
+ 1
3658
+ ]
3659
+ ],
3660
+ "flatten_loops": [
3661
+ true
3662
+ ],
3663
+ "l2_groupings": [
3664
+ 1
3665
+ ],
3666
+ "range_unroll_factors": [
3667
+ 0
3668
+ ],
3669
+ "range_warp_specializes": [],
3670
+ "range_num_stages": [
3671
+ 0
3672
+ ],
3673
+ "range_multi_buffers": [
3674
+ null
3675
+ ],
3676
+ "range_flattens": [
3677
+ null
3678
+ ],
3679
+ "load_eviction_policies": [
3680
+ "",
3681
+ "",
3682
+ ""
3683
+ ],
3684
+ "num_warps": 8,
3685
+ "num_stages": 1,
3686
+ "indexing": [
3687
+ "pointer",
3688
+ "pointer",
3689
+ "pointer",
3690
+ "pointer"
3691
+ ],
3692
+ "pid_type": "flat"
3693
+ },
3694
+ "intermediate_4096_numtokens_80": {
3695
+ "block_sizes": [
3696
+ 64,
3697
+ 64
3698
+ ],
3699
+ "loop_orders": [
3700
+ [
3701
+ 0,
3702
+ 1
3703
+ ]
3704
+ ],
3705
+ "flatten_loops": [
3706
+ true
3707
+ ],
3708
+ "l2_groupings": [
3709
+ 1
3710
+ ],
3711
+ "range_unroll_factors": [
3712
+ 0
3713
+ ],
3714
+ "range_warp_specializes": [],
3715
+ "range_num_stages": [
3716
+ 0
3717
+ ],
3718
+ "range_multi_buffers": [
3719
+ null
3720
+ ],
3721
+ "range_flattens": [
3722
+ null
3723
+ ],
3724
+ "load_eviction_policies": [
3725
+ "",
3726
+ "",
3727
+ ""
3728
+ ],
3729
+ "num_warps": 8,
3730
+ "num_stages": 1,
3731
+ "indexing": [
3732
+ "pointer",
3733
+ "pointer",
3734
+ "pointer",
3735
+ "pointer"
3736
+ ],
3737
+ "pid_type": "flat"
3738
+ },
3739
+ "intermediate_8192_numtokens_80": {
3740
+ "block_sizes": [
3741
+ 4,
3742
+ 128
3743
+ ],
3744
+ "loop_orders": [
3745
+ [
3746
+ 0,
3747
+ 1
3748
+ ]
3749
+ ],
3750
+ "flatten_loops": [
3751
+ true
3752
+ ],
3753
+ "l2_groupings": [
3754
+ 4
3755
+ ],
3756
+ "range_unroll_factors": [
3757
+ 0
3758
+ ],
3759
+ "range_warp_specializes": [],
3760
+ "range_num_stages": [
3761
+ 0
3762
+ ],
3763
+ "range_multi_buffers": [
3764
+ null
3765
+ ],
3766
+ "range_flattens": [
3767
+ null
3768
+ ],
3769
+ "load_eviction_policies": [
3770
+ "last",
3771
+ "last",
3772
+ "last"
3773
+ ],
3774
+ "num_warps": 16,
3775
+ "num_stages": 2,
3776
+ "indexing": [
3777
+ "pointer",
3778
+ "pointer",
3779
+ "pointer",
3780
+ "pointer"
3781
+ ],
3782
+ "pid_type": "flat"
3783
+ },
3784
+ "intermediate_11008_numtokens_80": {
3785
+ "block_sizes": [
3786
+ 1,
3787
+ 4096
3788
+ ],
3789
+ "loop_orders": [
3790
+ [
3791
+ 0,
3792
+ 1
3793
+ ]
3794
+ ],
3795
+ "flatten_loops": [
3796
+ true
3797
+ ],
3798
+ "l2_groupings": [
3799
+ 32
3800
+ ],
3801
+ "range_unroll_factors": [
3802
+ 0
3803
+ ],
3804
+ "range_warp_specializes": [],
3805
+ "range_num_stages": [
3806
+ 0
3807
+ ],
3808
+ "range_multi_buffers": [
3809
+ null
3810
+ ],
3811
+ "range_flattens": [
3812
+ null
3813
+ ],
3814
+ "load_eviction_policies": [
3815
+ "last",
3816
+ "last",
3817
+ ""
3818
+ ],
3819
+ "num_warps": 16,
3820
+ "num_stages": 2,
3821
+ "indexing": [
3822
+ "pointer",
3823
+ "tensor_descriptor",
3824
+ "tensor_descriptor",
3825
+ "tensor_descriptor"
3826
+ ],
3827
+ "pid_type": "flat"
3828
+ },
3829
+ "intermediate_14336_numtokens_80": {
3830
+ "block_sizes": [
3831
+ 2,
3832
+ 2048
3833
+ ],
3834
+ "loop_orders": [
3835
+ [
3836
+ 0,
3837
+ 1
3838
+ ]
3839
+ ],
3840
+ "flatten_loops": [
3841
+ true
3842
+ ],
3843
+ "l2_groupings": [
3844
+ 4
3845
+ ],
3846
+ "range_unroll_factors": [
3847
+ 0
3848
+ ],
3849
+ "range_warp_specializes": [],
3850
+ "range_num_stages": [
3851
+ 0
3852
+ ],
3853
+ "range_multi_buffers": [
3854
+ null
3855
+ ],
3856
+ "range_flattens": [
3857
+ null
3858
+ ],
3859
+ "load_eviction_policies": [
3860
+ "last",
3861
+ "last",
3862
+ ""
3863
+ ],
3864
+ "num_warps": 4,
3865
+ "num_stages": 6,
3866
+ "indexing": [
3867
+ "tensor_descriptor",
3868
+ "pointer",
3869
+ "pointer",
3870
+ "tensor_descriptor"
3871
+ ],
3872
+ "pid_type": "flat"
3873
+ },
3874
+ "intermediate_2048_numtokens_88": {
3875
+ "block_sizes": [
3876
+ 64,
3877
+ 32
3878
+ ],
3879
+ "loop_orders": [
3880
+ [
3881
+ 0,
3882
+ 1
3883
+ ]
3884
+ ],
3885
+ "flatten_loops": [
3886
+ true
3887
+ ],
3888
+ "l2_groupings": [
3889
+ 1
3890
+ ],
3891
+ "range_unroll_factors": [
3892
+ 0
3893
+ ],
3894
+ "range_warp_specializes": [],
3895
+ "range_num_stages": [
3896
+ 0
3897
+ ],
3898
+ "range_multi_buffers": [
3899
+ null
3900
+ ],
3901
+ "range_flattens": [
3902
+ null
3903
+ ],
3904
+ "load_eviction_policies": [
3905
+ "",
3906
+ "",
3907
+ ""
3908
+ ],
3909
+ "num_warps": 8,
3910
+ "num_stages": 1,
3911
+ "indexing": [
3912
+ "pointer",
3913
+ "pointer",
3914
+ "pointer",
3915
+ "pointer"
3916
+ ],
3917
+ "pid_type": "flat"
3918
+ },
3919
+ "intermediate_2880_numtokens_88": {
3920
+ "block_sizes": [
3921
+ 8,
3922
+ 256
3923
+ ],
3924
+ "loop_orders": [
3925
+ [
3926
+ 0,
3927
+ 1
3928
+ ]
3929
+ ],
3930
+ "flatten_loops": [
3931
+ true
3932
+ ],
3933
+ "l2_groupings": [
3934
+ 1
3935
+ ],
3936
+ "range_unroll_factors": [
3937
+ 0
3938
+ ],
3939
+ "range_warp_specializes": [],
3940
+ "range_num_stages": [
3941
+ 0
3942
+ ],
3943
+ "range_multi_buffers": [
3944
+ null
3945
+ ],
3946
+ "range_flattens": [
3947
+ null
3948
+ ],
3949
+ "load_eviction_policies": [
3950
+ "",
3951
+ "",
3952
+ ""
3953
+ ],
3954
+ "num_warps": 8,
3955
+ "num_stages": 1,
3956
+ "indexing": [
3957
+ "pointer",
3958
+ "pointer",
3959
+ "pointer",
3960
+ "pointer"
3961
+ ],
3962
+ "pid_type": "flat"
3963
+ },
3964
+ "intermediate_4096_numtokens_88": {
3965
+ "block_sizes": [
3966
+ 64,
3967
+ 64
3968
+ ],
3969
+ "loop_orders": [
3970
+ [
3971
+ 0,
3972
+ 1
3973
+ ]
3974
+ ],
3975
+ "flatten_loops": [
3976
+ true
3977
+ ],
3978
+ "l2_groupings": [
3979
+ 1
3980
+ ],
3981
+ "range_unroll_factors": [
3982
+ 0
3983
+ ],
3984
+ "range_warp_specializes": [],
3985
+ "range_num_stages": [
3986
+ 0
3987
+ ],
3988
+ "range_multi_buffers": [
3989
+ null
3990
+ ],
3991
+ "range_flattens": [
3992
+ null
3993
+ ],
3994
+ "load_eviction_policies": [
3995
+ "",
3996
+ "",
3997
+ ""
3998
+ ],
3999
+ "num_warps": 8,
4000
+ "num_stages": 1,
4001
+ "indexing": [
4002
+ "pointer",
4003
+ "pointer",
4004
+ "pointer",
4005
+ "pointer"
4006
+ ],
4007
+ "pid_type": "flat"
4008
+ },
4009
+ "intermediate_8192_numtokens_88": {
4010
+ "block_sizes": [
4011
+ 64,
4012
+ 128
4013
+ ],
4014
+ "loop_orders": [
4015
+ [
4016
+ 0,
4017
+ 1
4018
+ ]
4019
+ ],
4020
+ "flatten_loops": [
4021
+ true
4022
+ ],
4023
+ "l2_groupings": [
4024
+ 1
4025
+ ],
4026
+ "range_unroll_factors": [
4027
+ 0
4028
+ ],
4029
+ "range_warp_specializes": [],
4030
+ "range_num_stages": [
4031
+ 0
4032
+ ],
4033
+ "range_multi_buffers": [
4034
+ null
4035
+ ],
4036
+ "range_flattens": [
4037
+ null
4038
+ ],
4039
+ "load_eviction_policies": [
4040
+ "",
4041
+ "",
4042
+ ""
4043
+ ],
4044
+ "num_warps": 8,
4045
+ "num_stages": 1,
4046
+ "indexing": [
4047
+ "pointer",
4048
+ "pointer",
4049
+ "pointer",
4050
+ "pointer"
4051
+ ],
4052
+ "pid_type": "flat"
4053
+ },
4054
+ "intermediate_11008_numtokens_88": {
4055
+ "block_sizes": [
4056
+ 16,
4057
+ 2048
4058
+ ],
4059
+ "loop_orders": [
4060
+ [
4061
+ 0,
4062
+ 1
4063
+ ]
4064
+ ],
4065
+ "flatten_loops": [
4066
+ true
4067
+ ],
4068
+ "l2_groupings": [
4069
+ 2
4070
+ ],
4071
+ "range_unroll_factors": [
4072
+ 0
4073
+ ],
4074
+ "range_warp_specializes": [],
4075
+ "range_num_stages": [
4076
+ 0
4077
+ ],
4078
+ "range_multi_buffers": [
4079
+ null
4080
+ ],
4081
+ "range_flattens": [
4082
+ null
4083
+ ],
4084
+ "load_eviction_policies": [
4085
+ "first",
4086
+ "first",
4087
+ ""
4088
+ ],
4089
+ "num_warps": 32,
4090
+ "num_stages": 2,
4091
+ "indexing": [
4092
+ "tensor_descriptor",
4093
+ "tensor_descriptor",
4094
+ "pointer",
4095
+ "pointer"
4096
+ ],
4097
+ "pid_type": "flat"
4098
+ },
4099
+ "intermediate_14336_numtokens_88": {
4100
+ "block_sizes": [
4101
+ 4,
4102
+ 512
4103
+ ],
4104
+ "loop_orders": [
4105
+ [
4106
+ 0,
4107
+ 1
4108
+ ]
4109
+ ],
4110
+ "flatten_loops": [
4111
+ true
4112
+ ],
4113
+ "l2_groupings": [
4114
+ 16
4115
+ ],
4116
+ "range_unroll_factors": [
4117
+ 0
4118
+ ],
4119
+ "range_warp_specializes": [],
4120
+ "range_num_stages": [
4121
+ 0
4122
+ ],
4123
+ "range_multi_buffers": [
4124
+ null
4125
+ ],
4126
+ "range_flattens": [
4127
+ null
4128
+ ],
4129
+ "load_eviction_policies": [
4130
+ "",
4131
+ "",
4132
+ ""
4133
+ ],
4134
+ "num_warps": 32,
4135
+ "num_stages": 5,
4136
+ "indexing": [
4137
+ "pointer",
4138
+ "pointer",
4139
+ "pointer",
4140
+ "pointer"
4141
+ ],
4142
+ "pid_type": "flat"
4143
+ },
4144
+ "intermediate_2048_numtokens_96": {
4145
+ "block_sizes": [
4146
+ 64,
4147
+ 32
4148
+ ],
4149
+ "loop_orders": [
4150
+ [
4151
+ 0,
4152
+ 1
4153
+ ]
4154
+ ],
4155
+ "flatten_loops": [
4156
+ true
4157
+ ],
4158
+ "l2_groupings": [
4159
+ 1
4160
+ ],
4161
+ "range_unroll_factors": [
4162
+ 0
4163
+ ],
4164
+ "range_warp_specializes": [],
4165
+ "range_num_stages": [
4166
+ 0
4167
+ ],
4168
+ "range_multi_buffers": [
4169
+ null
4170
+ ],
4171
+ "range_flattens": [
4172
+ null
4173
+ ],
4174
+ "load_eviction_policies": [
4175
+ "",
4176
+ "",
4177
+ ""
4178
+ ],
4179
+ "num_warps": 8,
4180
+ "num_stages": 1,
4181
+ "indexing": [
4182
+ "pointer",
4183
+ "pointer",
4184
+ "pointer",
4185
+ "pointer"
4186
+ ],
4187
+ "pid_type": "flat"
4188
+ },
4189
+ "intermediate_2880_numtokens_96": {
4190
+ "block_sizes": [
4191
+ 32,
4192
+ 64
4193
+ ],
4194
+ "loop_orders": [
4195
+ [
4196
+ 0,
4197
+ 1
4198
+ ]
4199
+ ],
4200
+ "flatten_loops": [
4201
+ true
4202
+ ],
4203
+ "l2_groupings": [
4204
+ 1
4205
+ ],
4206
+ "range_unroll_factors": [
4207
+ 0
4208
+ ],
4209
+ "range_warp_specializes": [],
4210
+ "range_num_stages": [
4211
+ 0
4212
+ ],
4213
+ "range_multi_buffers": [
4214
+ null
4215
+ ],
4216
+ "range_flattens": [
4217
+ null
4218
+ ],
4219
+ "load_eviction_policies": [
4220
+ "",
4221
+ "",
4222
+ ""
4223
+ ],
4224
+ "num_warps": 8,
4225
+ "num_stages": 1,
4226
+ "indexing": [
4227
+ "pointer",
4228
+ "pointer",
4229
+ "pointer",
4230
+ "pointer"
4231
+ ],
4232
+ "pid_type": "flat"
4233
+ },
4234
+ "intermediate_4096_numtokens_96": {
4235
+ "block_sizes": [
4236
+ 1,
4237
+ 4096
4238
+ ],
4239
+ "loop_orders": [
4240
+ [
4241
+ 1,
4242
+ 0
4243
+ ]
4244
+ ],
4245
+ "flatten_loops": [
4246
+ false
4247
+ ],
4248
+ "l2_groupings": [
4249
+ 64
4250
+ ],
4251
+ "range_unroll_factors": [
4252
+ 0
4253
+ ],
4254
+ "range_warp_specializes": [],
4255
+ "range_num_stages": [
4256
+ 0
4257
+ ],
4258
+ "range_multi_buffers": [
4259
+ null
4260
+ ],
4261
+ "range_flattens": [
4262
+ null
4263
+ ],
4264
+ "load_eviction_policies": [
4265
+ "first",
4266
+ "last",
4267
+ "last"
4268
+ ],
4269
+ "num_warps": 4,
4270
+ "num_stages": 6,
4271
+ "indexing": [
4272
+ "pointer",
4273
+ "pointer",
4274
+ "tensor_descriptor",
4275
+ "pointer"
4276
+ ],
4277
+ "pid_type": "flat"
4278
+ },
4279
+ "intermediate_8192_numtokens_96": {
4280
+ "block_sizes": [
4281
+ 64,
4282
+ 128
4283
+ ],
4284
+ "loop_orders": [
4285
+ [
4286
+ 0,
4287
+ 1
4288
+ ]
4289
+ ],
4290
+ "flatten_loops": [
4291
+ true
4292
+ ],
4293
+ "l2_groupings": [
4294
+ 1
4295
+ ],
4296
+ "range_unroll_factors": [
4297
+ 0
4298
+ ],
4299
+ "range_warp_specializes": [],
4300
+ "range_num_stages": [
4301
+ 0
4302
+ ],
4303
+ "range_multi_buffers": [
4304
+ null
4305
+ ],
4306
+ "range_flattens": [
4307
+ null
4308
+ ],
4309
+ "load_eviction_policies": [
4310
+ "",
4311
+ "",
4312
+ ""
4313
+ ],
4314
+ "num_warps": 8,
4315
+ "num_stages": 1,
4316
+ "indexing": [
4317
+ "pointer",
4318
+ "pointer",
4319
+ "pointer",
4320
+ "pointer"
4321
+ ],
4322
+ "pid_type": "flat"
4323
+ },
4324
+ "intermediate_11008_numtokens_96": {
4325
+ "block_sizes": [
4326
+ 1,
4327
+ 2048
4328
+ ],
4329
+ "loop_orders": [
4330
+ [
4331
+ 1,
4332
+ 0
4333
+ ]
4334
+ ],
4335
+ "flatten_loops": [
4336
+ false
4337
+ ],
4338
+ "l2_groupings": [
4339
+ 64
4340
+ ],
4341
+ "range_unroll_factors": [
4342
+ 0
4343
+ ],
4344
+ "range_warp_specializes": [],
4345
+ "range_num_stages": [
4346
+ 0
4347
+ ],
4348
+ "range_multi_buffers": [
4349
+ null
4350
+ ],
4351
+ "range_flattens": [
4352
+ null
4353
+ ],
4354
+ "load_eviction_policies": [
4355
+ "first",
4356
+ "",
4357
+ ""
4358
+ ],
4359
+ "num_warps": 32,
4360
+ "num_stages": 2,
4361
+ "indexing": [
4362
+ "tensor_descriptor",
4363
+ "pointer",
4364
+ "pointer",
4365
+ "tensor_descriptor"
4366
+ ],
4367
+ "pid_type": "flat"
4368
+ },
4369
+ "intermediate_14336_numtokens_96": {
4370
+ "block_sizes": [
4371
+ 4,
4372
+ 4096
4373
+ ],
4374
+ "loop_orders": [
4375
+ [
4376
+ 0,
4377
+ 1
4378
+ ]
4379
+ ],
4380
+ "flatten_loops": [
4381
+ false
4382
+ ],
4383
+ "l2_groupings": [
4384
+ 32
4385
+ ],
4386
+ "range_unroll_factors": [
4387
+ 0
4388
+ ],
4389
+ "range_warp_specializes": [],
4390
+ "range_num_stages": [
4391
+ 0
4392
+ ],
4393
+ "range_multi_buffers": [
4394
+ null
4395
+ ],
4396
+ "range_flattens": [
4397
+ null
4398
+ ],
4399
+ "load_eviction_policies": [
4400
+ "last",
4401
+ "first",
4402
+ ""
4403
+ ],
4404
+ "num_warps": 4,
4405
+ "num_stages": 1,
4406
+ "indexing": [
4407
+ "tensor_descriptor",
4408
+ "tensor_descriptor",
4409
+ "pointer",
4410
+ "tensor_descriptor"
4411
+ ],
4412
+ "pid_type": "flat"
4413
+ },
4414
+ "intermediate_2048_numtokens_104": {
4415
+ "block_sizes": [
4416
+ 64,
4417
+ 32
4418
+ ],
4419
+ "loop_orders": [
4420
+ [
4421
+ 0,
4422
+ 1
4423
+ ]
4424
+ ],
4425
+ "flatten_loops": [
4426
+ true
4427
+ ],
4428
+ "l2_groupings": [
4429
+ 1
4430
+ ],
4431
+ "range_unroll_factors": [
4432
+ 0
4433
+ ],
4434
+ "range_warp_specializes": [],
4435
+ "range_num_stages": [
4436
+ 0
4437
+ ],
4438
+ "range_multi_buffers": [
4439
+ null
4440
+ ],
4441
+ "range_flattens": [
4442
+ null
4443
+ ],
4444
+ "load_eviction_policies": [
4445
+ "",
4446
+ "",
4447
+ ""
4448
+ ],
4449
+ "num_warps": 8,
4450
+ "num_stages": 1,
4451
+ "indexing": [
4452
+ "pointer",
4453
+ "pointer",
4454
+ "pointer",
4455
+ "pointer"
4456
+ ],
4457
+ "pid_type": "flat"
4458
+ },
4459
+ "intermediate_2880_numtokens_104": {
4460
+ "block_sizes": [
4461
+ 8,
4462
+ 256
4463
+ ],
4464
+ "loop_orders": [
4465
+ [
4466
+ 0,
4467
+ 1
4468
+ ]
4469
+ ],
4470
+ "flatten_loops": [
4471
+ true
4472
+ ],
4473
+ "l2_groupings": [
4474
+ 1
4475
+ ],
4476
+ "range_unroll_factors": [
4477
+ 0
4478
+ ],
4479
+ "range_warp_specializes": [],
4480
+ "range_num_stages": [
4481
+ 0
4482
+ ],
4483
+ "range_multi_buffers": [
4484
+ null
4485
+ ],
4486
+ "range_flattens": [
4487
+ null
4488
+ ],
4489
+ "load_eviction_policies": [
4490
+ "",
4491
+ "",
4492
+ ""
4493
+ ],
4494
+ "num_warps": 8,
4495
+ "num_stages": 1,
4496
+ "indexing": [
4497
+ "pointer",
4498
+ "pointer",
4499
+ "pointer",
4500
+ "pointer"
4501
+ ],
4502
+ "pid_type": "flat"
4503
+ },
4504
+ "intermediate_4096_numtokens_104": {
4505
+ "block_sizes": [
4506
+ 64,
4507
+ 64
4508
+ ],
4509
+ "loop_orders": [
4510
+ [
4511
+ 0,
4512
+ 1
4513
+ ]
4514
+ ],
4515
+ "flatten_loops": [
4516
+ true
4517
+ ],
4518
+ "l2_groupings": [
4519
+ 1
4520
+ ],
4521
+ "range_unroll_factors": [
4522
+ 0
4523
+ ],
4524
+ "range_warp_specializes": [],
4525
+ "range_num_stages": [
4526
+ 0
4527
+ ],
4528
+ "range_multi_buffers": [
4529
+ null
4530
+ ],
4531
+ "range_flattens": [
4532
+ null
4533
+ ],
4534
+ "load_eviction_policies": [
4535
+ "",
4536
+ "",
4537
+ ""
4538
+ ],
4539
+ "num_warps": 8,
4540
+ "num_stages": 1,
4541
+ "indexing": [
4542
+ "pointer",
4543
+ "pointer",
4544
+ "pointer",
4545
+ "pointer"
4546
+ ],
4547
+ "pid_type": "flat"
4548
+ },
4549
+ "intermediate_8192_numtokens_104": {
4550
+ "block_sizes": [
4551
+ 64,
4552
+ 128
4553
+ ],
4554
+ "loop_orders": [
4555
+ [
4556
+ 0,
4557
+ 1
4558
+ ]
4559
+ ],
4560
+ "flatten_loops": [
4561
+ true
4562
+ ],
4563
+ "l2_groupings": [
4564
+ 1
4565
+ ],
4566
+ "range_unroll_factors": [
4567
+ 0
4568
+ ],
4569
+ "range_warp_specializes": [],
4570
+ "range_num_stages": [
4571
+ 0
4572
+ ],
4573
+ "range_multi_buffers": [
4574
+ null
4575
+ ],
4576
+ "range_flattens": [
4577
+ null
4578
+ ],
4579
+ "load_eviction_policies": [
4580
+ "",
4581
+ "",
4582
+ ""
4583
+ ],
4584
+ "num_warps": 8,
4585
+ "num_stages": 1,
4586
+ "indexing": [
4587
+ "pointer",
4588
+ "pointer",
4589
+ "pointer",
4590
+ "pointer"
4591
+ ],
4592
+ "pid_type": "flat"
4593
+ },
4594
+ "intermediate_11008_numtokens_104": {
4595
+ "block_sizes": [
4596
+ 2,
4597
+ 8192
4598
+ ],
4599
+ "loop_orders": [
4600
+ [
4601
+ 0,
4602
+ 1
4603
+ ]
4604
+ ],
4605
+ "flatten_loops": [
4606
+ true
4607
+ ],
4608
+ "l2_groupings": [
4609
+ 1
4610
+ ],
4611
+ "range_unroll_factors": [
4612
+ 0
4613
+ ],
4614
+ "range_warp_specializes": [],
4615
+ "range_num_stages": [
4616
+ 0
4617
+ ],
4618
+ "range_multi_buffers": [
4619
+ null
4620
+ ],
4621
+ "range_flattens": [
4622
+ null
4623
+ ],
4624
+ "load_eviction_policies": [
4625
+ "",
4626
+ "last",
4627
+ "first"
4628
+ ],
4629
+ "num_warps": 8,
4630
+ "num_stages": 1,
4631
+ "indexing": [
4632
+ "pointer",
4633
+ "tensor_descriptor",
4634
+ "tensor_descriptor",
4635
+ "pointer"
4636
+ ],
4637
+ "pid_type": "flat"
4638
+ },
4639
+ "intermediate_14336_numtokens_104": {
4640
+ "block_sizes": [
4641
+ 8,
4642
+ 512
4643
+ ],
4644
+ "loop_orders": [
4645
+ [
4646
+ 0,
4647
+ 1
4648
+ ]
4649
+ ],
4650
+ "flatten_loops": [
4651
+ true
4652
+ ],
4653
+ "l2_groupings": [
4654
+ 2
4655
+ ],
4656
+ "range_unroll_factors": [
4657
+ 0
4658
+ ],
4659
+ "range_warp_specializes": [],
4660
+ "range_num_stages": [
4661
+ 0
4662
+ ],
4663
+ "range_multi_buffers": [
4664
+ null
4665
+ ],
4666
+ "range_flattens": [
4667
+ null
4668
+ ],
4669
+ "load_eviction_policies": [
4670
+ "last",
4671
+ "last",
4672
+ ""
4673
+ ],
4674
+ "num_warps": 8,
4675
+ "num_stages": 1,
4676
+ "indexing": [
4677
+ "pointer",
4678
+ "tensor_descriptor",
4679
+ "pointer",
4680
+ "tensor_descriptor"
4681
+ ],
4682
+ "pid_type": "flat"
4683
+ },
4684
+ "intermediate_2048_numtokens_112": {
4685
+ "block_sizes": [
4686
+ 64,
4687
+ 32
4688
+ ],
4689
+ "loop_orders": [
4690
+ [
4691
+ 0,
4692
+ 1
4693
+ ]
4694
+ ],
4695
+ "flatten_loops": [
4696
+ true
4697
+ ],
4698
+ "l2_groupings": [
4699
+ 1
4700
+ ],
4701
+ "range_unroll_factors": [
4702
+ 0
4703
+ ],
4704
+ "range_warp_specializes": [],
4705
+ "range_num_stages": [
4706
+ 0
4707
+ ],
4708
+ "range_multi_buffers": [
4709
+ null
4710
+ ],
4711
+ "range_flattens": [
4712
+ null
4713
+ ],
4714
+ "load_eviction_policies": [
4715
+ "",
4716
+ "",
4717
+ ""
4718
+ ],
4719
+ "num_warps": 8,
4720
+ "num_stages": 1,
4721
+ "indexing": [
4722
+ "pointer",
4723
+ "pointer",
4724
+ "pointer",
4725
+ "pointer"
4726
+ ],
4727
+ "pid_type": "flat"
4728
+ },
4729
+ "intermediate_2880_numtokens_112": {
4730
+ "block_sizes": [
4731
+ 2,
4732
+ 2048
4733
+ ],
4734
+ "loop_orders": [
4735
+ [
4736
+ 0,
4737
+ 1
4738
+ ]
4739
+ ],
4740
+ "flatten_loops": [
4741
+ true
4742
+ ],
4743
+ "l2_groupings": [
4744
+ 1
4745
+ ],
4746
+ "range_unroll_factors": [
4747
+ 0
4748
+ ],
4749
+ "range_warp_specializes": [],
4750
+ "range_num_stages": [
4751
+ 0
4752
+ ],
4753
+ "range_multi_buffers": [
4754
+ null
4755
+ ],
4756
+ "range_flattens": [
4757
+ null
4758
+ ],
4759
+ "load_eviction_policies": [
4760
+ "",
4761
+ "",
4762
+ ""
4763
+ ],
4764
+ "num_warps": 8,
4765
+ "num_stages": 1,
4766
+ "indexing": [
4767
+ "pointer",
4768
+ "pointer",
4769
+ "pointer",
4770
+ "pointer"
4771
+ ],
4772
+ "pid_type": "flat"
4773
+ },
4774
+ "intermediate_4096_numtokens_112": {
4775
+ "block_sizes": [
4776
+ 64,
4777
+ 64
4778
+ ],
4779
+ "loop_orders": [
4780
+ [
4781
+ 0,
4782
+ 1
4783
+ ]
4784
+ ],
4785
+ "flatten_loops": [
4786
+ true
4787
+ ],
4788
+ "l2_groupings": [
4789
+ 1
4790
+ ],
4791
+ "range_unroll_factors": [
4792
+ 0
4793
+ ],
4794
+ "range_warp_specializes": [],
4795
+ "range_num_stages": [
4796
+ 0
4797
+ ],
4798
+ "range_multi_buffers": [
4799
+ null
4800
+ ],
4801
+ "range_flattens": [
4802
+ null
4803
+ ],
4804
+ "load_eviction_policies": [
4805
+ "",
4806
+ "",
4807
+ ""
4808
+ ],
4809
+ "num_warps": 8,
4810
+ "num_stages": 1,
4811
+ "indexing": [
4812
+ "pointer",
4813
+ "pointer",
4814
+ "pointer",
4815
+ "pointer"
4816
+ ],
4817
+ "pid_type": "flat"
4818
+ },
4819
+ "intermediate_8192_numtokens_112": {
4820
+ "block_sizes": [
4821
+ 4,
4822
+ 512
4823
+ ],
4824
+ "loop_orders": [
4825
+ [
4826
+ 0,
4827
+ 1
4828
+ ]
4829
+ ],
4830
+ "flatten_loops": [
4831
+ true
4832
+ ],
4833
+ "l2_groupings": [
4834
+ 16
4835
+ ],
4836
+ "range_unroll_factors": [
4837
+ 0
4838
+ ],
4839
+ "range_warp_specializes": [],
4840
+ "range_num_stages": [
4841
+ 0
4842
+ ],
4843
+ "range_multi_buffers": [
4844
+ null
4845
+ ],
4846
+ "range_flattens": [
4847
+ null
4848
+ ],
4849
+ "load_eviction_policies": [
4850
+ "first",
4851
+ "last",
4852
+ "last"
4853
+ ],
4854
+ "num_warps": 16,
4855
+ "num_stages": 3,
4856
+ "indexing": [
4857
+ "pointer",
4858
+ "pointer",
4859
+ "tensor_descriptor",
4860
+ "pointer"
4861
+ ],
4862
+ "pid_type": "flat"
4863
+ },
4864
+ "intermediate_11008_numtokens_112": {
4865
+ "block_sizes": [
4866
+ 4,
4867
+ 1024
4868
+ ],
4869
+ "loop_orders": [
4870
+ [
4871
+ 0,
4872
+ 1
4873
+ ]
4874
+ ],
4875
+ "flatten_loops": [
4876
+ true
4877
+ ],
4878
+ "l2_groupings": [
4879
+ 2
4880
+ ],
4881
+ "range_unroll_factors": [
4882
+ 0
4883
+ ],
4884
+ "range_warp_specializes": [],
4885
+ "range_num_stages": [
4886
+ 0
4887
+ ],
4888
+ "range_multi_buffers": [
4889
+ null
4890
+ ],
4891
+ "range_flattens": [
4892
+ null
4893
+ ],
4894
+ "load_eviction_policies": [
4895
+ "",
4896
+ "last",
4897
+ "last"
4898
+ ],
4899
+ "num_warps": 32,
4900
+ "num_stages": 3,
4901
+ "indexing": [
4902
+ "pointer",
4903
+ "pointer",
4904
+ "tensor_descriptor",
4905
+ "tensor_descriptor"
4906
+ ],
4907
+ "pid_type": "flat"
4908
+ },
4909
+ "intermediate_14336_numtokens_112": {
4910
+ "block_sizes": [
4911
+ 64,
4912
+ 256
4913
+ ],
4914
+ "loop_orders": [
4915
+ [
4916
+ 0,
4917
+ 1
4918
+ ]
4919
+ ],
4920
+ "flatten_loops": [
4921
+ true
4922
+ ],
4923
+ "l2_groupings": [
4924
+ 1
4925
+ ],
4926
+ "range_unroll_factors": [
4927
+ 0
4928
+ ],
4929
+ "range_warp_specializes": [],
4930
+ "range_num_stages": [
4931
+ 0
4932
+ ],
4933
+ "range_multi_buffers": [
4934
+ null
4935
+ ],
4936
+ "range_flattens": [
4937
+ null
4938
+ ],
4939
+ "load_eviction_policies": [
4940
+ "",
4941
+ "",
4942
+ ""
4943
+ ],
4944
+ "num_warps": 8,
4945
+ "num_stages": 1,
4946
+ "indexing": [
4947
+ "pointer",
4948
+ "pointer",
4949
+ "pointer",
4950
+ "pointer"
4951
+ ],
4952
+ "pid_type": "flat"
4953
+ },
4954
+ "intermediate_2048_numtokens_120": {
4955
+ "block_sizes": [
4956
+ 8,
4957
+ 256
4958
+ ],
4959
+ "loop_orders": [
4960
+ [
4961
+ 0,
4962
+ 1
4963
+ ]
4964
+ ],
4965
+ "flatten_loops": [
4966
+ true
4967
+ ],
4968
+ "l2_groupings": [
4969
+ 16
4970
+ ],
4971
+ "range_unroll_factors": [
4972
+ 0
4973
+ ],
4974
+ "range_warp_specializes": [],
4975
+ "range_num_stages": [
4976
+ 0
4977
+ ],
4978
+ "range_multi_buffers": [
4979
+ null
4980
+ ],
4981
+ "range_flattens": [
4982
+ null
4983
+ ],
4984
+ "load_eviction_policies": [
4985
+ "first",
4986
+ "last",
4987
+ ""
4988
+ ],
4989
+ "num_warps": 4,
4990
+ "num_stages": 6,
4991
+ "indexing": [
4992
+ "tensor_descriptor",
4993
+ "tensor_descriptor",
4994
+ "pointer",
4995
+ "tensor_descriptor"
4996
+ ],
4997
+ "pid_type": "flat"
4998
+ },
4999
+ "intermediate_2880_numtokens_120": {
This file is too large to show in full.
vllm/kernels/helion/configs/silu_mul_fp8/nvidia_h200.json
new
+13866
@@ -0,0 +1,13866 @@
1
+{
2
+ "intermediate_2048_numtokens_256": {
3
+ "block_sizes": [
4
+ 64,
5
+ 32
6
+ ],
7
+ "loop_orders": [
8
+ [
9
+ 0,
10
+ 1
11
+ ]
12
+ ],
13
+ "flatten_loops": [
14
+ true
15
+ ],
16
+ "l2_groupings": [
17
+ 1
18
+ ],
19
+ "range_unroll_factors": [
20
+ 0
21
+ ],
22
+ "range_warp_specializes": [],
23
+ "range_num_stages": [
24
+ 0
25
+ ],
26
+ "range_multi_buffers": [
27
+ null
28
+ ],
29
+ "range_flattens": [
30
+ null
31
+ ],
32
+ "load_eviction_policies": [
33
+ "",
34
+ "",
35
+ ""
36
+ ],
37
+ "num_warps": 8,
38
+ "num_stages": 1,
39
+ "indexing": [
40
+ "pointer",
41
+ "pointer",
42
+ "pointer",
43
+ "pointer"
44
+ ],
45
+ "pid_type": "flat"
46
+ },
47
+ "intermediate_4096_numtokens_256": {
48
+ "block_sizes": [
49
+ 256,
50
+ 32
51
+ ],
52
+ "loop_orders": [
53
+ [
54
+ 0,
55
+ 1
56
+ ]
57
+ ],
58
+ "flatten_loops": [
59
+ true
60
+ ],
61
+ "l2_groupings": [
62
+ 1
63
+ ],
64
+ "range_unroll_factors": [
65
+ 0
66
+ ],
67
+ "range_warp_specializes": [],
68
+ "range_num_stages": [
69
+ 0
70
+ ],
71
+ "range_multi_buffers": [
72
+ null
73
+ ],
74
+ "range_flattens": [
75
+ null
76
+ ],
77
+ "load_eviction_policies": [
78
+ "",
79
+ "",
80
+ ""
81
+ ],
82
+ "num_warps": 8,
83
+ "num_stages": 1,
84
+ "indexing": [
85
+ "pointer",
86
+ "pointer",
87
+ "pointer",
88
+ "pointer"
89
+ ],
90
+ "pid_type": "flat"
91
+ },
92
+ "default": {
93
+ "block_sizes": [
94
+ 1,
95
+ 512
96
+ ],
97
+ "loop_orders": [
98
+ [
99
+ 1,
100
+ 0
101
+ ]
102
+ ],
103
+ "flatten_loops": [
104
+ false
105
+ ],
106
+ "l2_groupings": [
107
+ 4
108
+ ],
109
+ "range_unroll_factors": [
110
+ 0
111
+ ],
112
+ "range_warp_specializes": [],
113
+ "range_num_stages": [
114
+ 0
115
+ ],
116
+ "range_multi_buffers": [
117
+ null
118
+ ],
119
+ "range_flattens": [
120
+ null
121
+ ],
122
+ "load_eviction_policies": [
123
+ "",
124
+ "first",
125
+ ""
126
+ ],
127
+ "num_warps": 8,
128
+ "num_stages": 2,
129
+ "indexing": [
130
+ "tensor_descriptor",
131
+ "tensor_descriptor",
132
+ "tensor_descriptor",
133
+ "pointer"
134
+ ],
135
+ "pid_type": "flat"
136
+ },
137
+ "intermediate_8192_numtokens_256": {
138
+ "block_sizes": [
139
+ 256,
140
+ 64
141
+ ],
142
+ "loop_orders": [
143
+ [
144
+ 0,
145
+ 1
146
+ ]
147
+ ],
148
+ "flatten_loops": [
149
+ true
150
+ ],
151
+ "l2_groupings": [
152
+ 1
153
+ ],
154
+ "range_unroll_factors": [
155
+ 0
156
+ ],
157
+ "range_warp_specializes": [],
158
+ "range_num_stages": [
159
+ 0
160
+ ],
161
+ "range_multi_buffers": [
162
+ null
163
+ ],
164
+ "range_flattens": [
165
+ null
166
+ ],
167
+ "load_eviction_policies": [
168
+ "",
169
+ "",
170
+ ""
171
+ ],
172
+ "num_warps": 8,
173
+ "num_stages": 1,
174
+ "indexing": [
175
+ "pointer",
176
+ "pointer",
177
+ "pointer",
178
+ "pointer"
179
+ ],
180
+ "pid_type": "flat"
181
+ },
182
+ "intermediate_14336_numtokens_256": {
183
+ "block_sizes": [
184
+ 8,
185
+ 4096
186
+ ],
187
+ "loop_orders": [
188
+ [
189
+ 0,
190
+ 1
191
+ ]
192
+ ],
193
+ "flatten_loops": [
194
+ true
195
+ ],
196
+ "l2_groupings": [
197
+ 1
198
+ ],
199
+ "range_unroll_factors": [
200
+ 0
201
+ ],
202
+ "range_warp_specializes": [],
203
+ "range_num_stages": [
204
+ 0
205
+ ],
206
+ "range_multi_buffers": [
207
+ null
208
+ ],
209
+ "range_flattens": [
210
+ null
211
+ ],
212
+ "load_eviction_policies": [
213
+ "",
214
+ "",
215
+ ""
216
+ ],
217
+ "num_warps": 8,
218
+ "num_stages": 1,
219
+ "indexing": [
220
+ "pointer",
221
+ "pointer",
222
+ "pointer",
223
+ "pointer"
224
+ ],
225
+ "pid_type": "flat"
226
+ },
227
+ "intermediate_11008_numtokens_256": {
228
+ "block_sizes": [
229
+ 64,
230
+ 32
231
+ ],
232
+ "loop_orders": [
233
+ [
234
+ 0,
235
+ 1
236
+ ]
237
+ ],
238
+ "flatten_loops": [
239
+ true
240
+ ],
241
+ "l2_groupings": [
242
+ 1
243
+ ],
244
+ "range_unroll_factors": [
245
+ 0
246
+ ],
247
+ "range_warp_specializes": [],
248
+ "range_num_stages": [
249
+ 0
250
+ ],
251
+ "range_multi_buffers": [
252
+ null
253
+ ],
254
+ "range_flattens": [
255
+ null
256
+ ],
257
+ "load_eviction_policies": [
258
+ "",
259
+ "",
260
+ ""
261
+ ],
262
+ "num_warps": 2,
263
+ "num_stages": 1,
264
+ "indexing": [
265
+ "pointer",
266
+ "pointer",
267
+ "pointer",
268
+ "pointer"
269
+ ],
270
+ "pid_type": "flat"
271
+ },
272
+ "intermediate_7688_numtokens_256": {
273
+ "block_sizes": [
274
+ 32,
275
+ 512
276
+ ],
277
+ "loop_orders": [
278
+ [
279
+ 0,
280
+ 1
281
+ ]
282
+ ],
283
+ "flatten_loops": [
284
+ true
285
+ ],
286
+ "l2_groupings": [
287
+ 1
288
+ ],
289
+ "range_unroll_factors": [
290
+ 0
291
+ ],
292
+ "range_warp_specializes": [],
293
+ "range_num_stages": [
294
+ 0
295
+ ],
296
+ "range_multi_buffers": [
297
+ null
298
+ ],
299
+ "range_flattens": [
300
+ null
301
+ ],
302
+ "load_eviction_policies": [
303
+ "",
304
+ "",
305
+ ""
306
+ ],
307
+ "num_warps": 8,
308
+ "num_stages": 1,
309
+ "indexing": [
310
+ "pointer",
311
+ "pointer",
312
+ "pointer",
313
+ "pointer"
314
+ ],
315
+ "pid_type": "flat"
316
+ },
317
+ "intermediate_2880_numtokens_256": {
318
+ "block_sizes": [
319
+ 32,
320
+ 64
321
+ ],
322
+ "loop_orders": [
323
+ [
324
+ 0,
325
+ 1
326
+ ]
327
+ ],
328
+ "flatten_loops": [
329
+ true
330
+ ],
331
+ "l2_groupings": [
332
+ 2
333
+ ],
334
+ "range_unroll_factors": [
335
+ 0
336
+ ],
337
+ "range_warp_specializes": [],
338
+ "range_num_stages": [
339
+ 0
340
+ ],
341
+ "range_multi_buffers": [
342
+ null
343
+ ],
344
+ "range_flattens": [
345
+ null
346
+ ],
347
+ "load_eviction_policies": [
348
+ "",
349
+ "",
350
+ ""
351
+ ],
352
+ "num_warps": 8,
353
+ "num_stages": 1,
354
+ "indexing": [
355
+ "pointer",
356
+ "pointer",
357
+ "pointer",
358
+ "pointer"
359
+ ],
360
+ "pid_type": "flat"
361
+ },
362
+ "intermediate_2048_numtokens_1": {
363
+ "block_sizes": [
364
+ 1,
365
+ 256
366
+ ],
367
+ "loop_orders": [
368
+ [
369
+ 0,
370
+ 1
371
+ ]
372
+ ],
373
+ "flatten_loops": [
374
+ true
375
+ ],
376
+ "l2_groupings": [
377
+ 1
378
+ ],
379
+ "range_unroll_factors": [
380
+ 0
381
+ ],
382
+ "range_warp_specializes": [],
383
+ "range_num_stages": [
384
+ 0
385
+ ],
386
+ "range_multi_buffers": [
387
+ null
388
+ ],
389
+ "range_flattens": [
390
+ null
391
+ ],
392
+ "load_eviction_policies": [
393
+ "",
394
+ "",
395
+ ""
396
+ ],
397
+ "num_warps": 2,
398
+ "num_stages": 1,
399
+ "indexing": [
400
+ "pointer",
401
+ "pointer",
402
+ "pointer",
403
+ "pointer"
404
+ ],
405
+ "pid_type": "flat"
406
+ },
407
+ "intermediate_2880_numtokens_1": {
408
+ "block_sizes": [
409
+ 1,
410
+ 128
411
+ ],
412
+ "loop_orders": [
413
+ [
414
+ 0,
415
+ 1
416
+ ]
417
+ ],
418
+ "flatten_loops": [
419
+ true
420
+ ],
421
+ "l2_groupings": [
422
+ 1
423
+ ],
424
+ "range_unroll_factors": [
425
+ 0
426
+ ],
427
+ "range_warp_specializes": [],
428
+ "range_num_stages": [
429
+ 0
430
+ ],
431
+ "range_multi_buffers": [
432
+ null
433
+ ],
434
+ "range_flattens": [
435
+ null
436
+ ],
437
+ "load_eviction_policies": [
438
+ "",
439
+ "",
440
+ ""
441
+ ],
442
+ "num_warps": 4,
443
+ "num_stages": 1,
444
+ "indexing": [
445
+ "pointer",
446
+ "pointer",
447
+ "pointer",
448
+ "pointer"
449
+ ],
450
+ "pid_type": "flat"
451
+ },
452
+ "intermediate_4096_numtokens_1": {
453
+ "block_sizes": [
454
+ 1,
455
+ 256
456
+ ],
457
+ "loop_orders": [
458
+ [
459
+ 0,
460
+ 1
461
+ ]
462
+ ],
463
+ "flatten_loops": [
464
+ true
465
+ ],
466
+ "l2_groupings": [
467
+ 1
468
+ ],
469
+ "range_unroll_factors": [
470
+ 0
471
+ ],
472
+ "range_warp_specializes": [],
473
+ "range_num_stages": [
474
+ 0
475
+ ],
476
+ "range_multi_buffers": [
477
+ null
478
+ ],
479
+ "range_flattens": [
480
+ null
481
+ ],
482
+ "load_eviction_policies": [
483
+ "",
484
+ "",
485
+ ""
486
+ ],
487
+ "num_warps": 2,
488
+ "num_stages": 1,
489
+ "indexing": [
490
+ "pointer",
491
+ "pointer",
492
+ "pointer",
493
+ "pointer"
494
+ ],
495
+ "pid_type": "flat"
496
+ },
497
+ "intermediate_8192_numtokens_1": {
498
+ "block_sizes": [
499
+ 1,
500
+ 256
501
+ ],
502
+ "loop_orders": [
503
+ [
504
+ 0,
505
+ 1
506
+ ]
507
+ ],
508
+ "flatten_loops": [
509
+ true
510
+ ],
511
+ "l2_groupings": [
512
+ 1
513
+ ],
514
+ "range_unroll_factors": [
515
+ 0
516
+ ],
517
+ "range_warp_specializes": [],
518
+ "range_num_stages": [
519
+ 0
520
+ ],
521
+ "range_multi_buffers": [
522
+ null
523
+ ],
524
+ "range_flattens": [
525
+ null
526
+ ],
527
+ "load_eviction_policies": [
528
+ "",
529
+ "",
530
+ ""
531
+ ],
532
+ "num_warps": 2,
533
+ "num_stages": 1,
534
+ "indexing": [
535
+ "pointer",
536
+ "pointer",
537
+ "pointer",
538
+ "pointer"
539
+ ],
540
+ "pid_type": "flat"
541
+ },
542
+ "intermediate_11008_numtokens_1": {
543
+ "block_sizes": [
544
+ 1,
545
+ 256
546
+ ],
547
+ "loop_orders": [
548
+ [
549
+ 0,
550
+ 1
551
+ ]
552
+ ],
553
+ "flatten_loops": [
554
+ true
555
+ ],
556
+ "l2_groupings": [
557
+ 1
558
+ ],
559
+ "range_unroll_factors": [
560
+ 0
561
+ ],
562
+ "range_warp_specializes": [],
563
+ "range_num_stages": [
564
+ 0
565
+ ],
566
+ "range_multi_buffers": [
567
+ null
568
+ ],
569
+ "range_flattens": [
570
+ null
571
+ ],
572
+ "load_eviction_policies": [
573
+ "",
574
+ "",
575
+ ""
576
+ ],
577
+ "num_warps": 2,
578
+ "num_stages": 1,
579
+ "indexing": [
580
+ "pointer",
581
+ "pointer",
582
+ "pointer",
583
+ "pointer"
584
+ ],
585
+ "pid_type": "flat"
586
+ },
587
+ "intermediate_14336_numtokens_1": {
588
+ "block_sizes": [
589
+ 1,
590
+ 256
591
+ ],
592
+ "loop_orders": [
593
+ [
594
+ 0,
595
+ 1
596
+ ]
597
+ ],
598
+ "flatten_loops": [
599
+ true
600
+ ],
601
+ "l2_groupings": [
602
+ 1
603
+ ],
604
+ "range_unroll_factors": [
605
+ 0
606
+ ],
607
+ "range_warp_specializes": [],
608
+ "range_num_stages": [
609
+ 0
610
+ ],
611
+ "range_multi_buffers": [
612
+ null
613
+ ],
614
+ "range_flattens": [
615
+ null
616
+ ],
617
+ "load_eviction_policies": [
618
+ "",
619
+ "",
620
+ ""
621
+ ],
622
+ "num_warps": 2,
623
+ "num_stages": 1,
624
+ "indexing": [
625
+ "pointer",
626
+ "pointer",
627
+ "pointer",
628
+ "pointer"
629
+ ],
630
+ "pid_type": "flat"
631
+ },
632
+ "intermediate_2048_numtokens_2": {
633
+ "block_sizes": [
634
+ 2,
635
+ 128
636
+ ],
637
+ "loop_orders": [
638
+ [
639
+ 0,
640
+ 1
641
+ ]
642
+ ],
643
+ "flatten_loops": [
644
+ true
645
+ ],
646
+ "l2_groupings": [
647
+ 1
648
+ ],
649
+ "range_unroll_factors": [
650
+ 0
651
+ ],
652
+ "range_warp_specializes": [],
653
+ "range_num_stages": [
654
+ 0
655
+ ],
656
+ "range_multi_buffers": [
657
+ null
658
+ ],
659
+ "range_flattens": [
660
+ null
661
+ ],
662
+ "load_eviction_policies": [
663
+ "",
664
+ "",
665
+ ""
666
+ ],
667
+ "num_warps": 2,
668
+ "num_stages": 1,
669
+ "indexing": [
670
+ "pointer",
671
+ "pointer",
672
+ "pointer",
673
+ "pointer"
674
+ ],
675
+ "pid_type": "flat"
676
+ },
677
+ "intermediate_2880_numtokens_2": {
678
+ "block_sizes": [
679
+ 2,
680
+ 256
681
+ ],
682
+ "loop_orders": [
683
+ [
684
+ 0,
685
+ 1
686
+ ]
687
+ ],
688
+ "flatten_loops": [
689
+ true
690
+ ],
691
+ "l2_groupings": [
692
+ 1
693
+ ],
694
+ "range_unroll_factors": [
695
+ 0
696
+ ],
697
+ "range_warp_specializes": [],
698
+ "range_num_stages": [
699
+ 0
700
+ ],
701
+ "range_multi_buffers": [
702
+ null
703
+ ],
704
+ "range_flattens": [
705
+ null
706
+ ],
707
+ "load_eviction_policies": [
708
+ "",
709
+ "",
710
+ ""
711
+ ],
712
+ "num_warps": 4,
713
+ "num_stages": 1,
714
+ "indexing": [
715
+ "pointer",
716
+ "pointer",
717
+ "pointer",
718
+ "pointer"
719
+ ],
720
+ "pid_type": "flat"
721
+ },
722
+ "intermediate_4096_numtokens_2": {
723
+ "block_sizes": [
724
+ 2,
725
+ 128
726
+ ],
727
+ "loop_orders": [
728
+ [
729
+ 0,
730
+ 1
731
+ ]
732
+ ],
733
+ "flatten_loops": [
734
+ true
735
+ ],
736
+ "l2_groupings": [
737
+ 1
738
+ ],
739
+ "range_unroll_factors": [
740
+ 0
741
+ ],
742
+ "range_warp_specializes": [],
743
+ "range_num_stages": [
744
+ 0
745
+ ],
746
+ "range_multi_buffers": [
747
+ null
748
+ ],
749
+ "range_flattens": [
750
+ null
751
+ ],
752
+ "load_eviction_policies": [
753
+ "",
754
+ "",
755
+ ""
756
+ ],
757
+ "num_warps": 2,
758
+ "num_stages": 1,
759
+ "indexing": [
760
+ "pointer",
761
+ "pointer",
762
+ "pointer",
763
+ "pointer"
764
+ ],
765
+ "pid_type": "flat"
766
+ },
767
+ "intermediate_8192_numtokens_2": {
768
+ "block_sizes": [
769
+ 2,
770
+ 128
771
+ ],
772
+ "loop_orders": [
773
+ [
774
+ 0,
775
+ 1
776
+ ]
777
+ ],
778
+ "flatten_loops": [
779
+ true
780
+ ],
781
+ "l2_groupings": [
782
+ 1
783
+ ],
784
+ "range_unroll_factors": [
785
+ 0
786
+ ],
787
+ "range_warp_specializes": [],
788
+ "range_num_stages": [
789
+ 0
790
+ ],
791
+ "range_multi_buffers": [
792
+ null
793
+ ],
794
+ "range_flattens": [
795
+ null
796
+ ],
797
+ "load_eviction_policies": [
798
+ "",
799
+ "",
800
+ ""
801
+ ],
802
+ "num_warps": 2,
803
+ "num_stages": 1,
804
+ "indexing": [
805
+ "pointer",
806
+ "pointer",
807
+ "pointer",
808
+ "pointer"
809
+ ],
810
+ "pid_type": "flat"
811
+ },
812
+ "intermediate_11008_numtokens_2": {
813
+ "block_sizes": [
814
+ 1,
815
+ 16384
816
+ ],
817
+ "loop_orders": [
818
+ [
819
+ 1,
820
+ 0
821
+ ]
822
+ ],
823
+ "flatten_loops": [
824
+ true
825
+ ],
826
+ "l2_groupings": [
827
+ 16
828
+ ],
829
+ "range_unroll_factors": [
830
+ 0
831
+ ],
832
+ "range_warp_specializes": [],
833
+ "range_num_stages": [
834
+ 0
835
+ ],
836
+ "range_multi_buffers": [
837
+ null
838
+ ],
839
+ "range_flattens": [
840
+ null
841
+ ],
842
+ "load_eviction_policies": [
843
+ "last",
844
+ "first",
845
+ "last"
846
+ ],
847
+ "num_warps": 32,
848
+ "num_stages": 3,
849
+ "indexing": [
850
+ "tensor_descriptor",
851
+ "pointer",
852
+ "pointer",
853
+ "tensor_descriptor"
854
+ ],
855
+ "pid_type": "xyz"
856
+ },
857
+ "intermediate_14336_numtokens_2": {
858
+ "block_sizes": [
859
+ 2,
860
+ 128
861
+ ],
862
+ "loop_orders": [
863
+ [
864
+ 0,
865
+ 1
866
+ ]
867
+ ],
868
+ "flatten_loops": [
869
+ true
870
+ ],
871
+ "l2_groupings": [
872
+ 1
873
+ ],
874
+ "range_unroll_factors": [
875
+ 0
876
+ ],
877
+ "range_warp_specializes": [],
878
+ "range_num_stages": [
879
+ 0
880
+ ],
881
+ "range_multi_buffers": [
882
+ null
883
+ ],
884
+ "range_flattens": [
885
+ null
886
+ ],
887
+ "load_eviction_policies": [
888
+ "",
889
+ "",
890
+ ""
891
+ ],
892
+ "num_warps": 2,
893
+ "num_stages": 1,
894
+ "indexing": [
895
+ "pointer",
896
+ "pointer",
897
+ "pointer",
898
+ "pointer"
899
+ ],
900
+ "pid_type": "flat"
901
+ },
902
+ "intermediate_2048_numtokens_4": {
903
+ "block_sizes": [
904
+ 4,
905
+ 64
906
+ ],
907
+ "loop_orders": [
908
+ [
909
+ 0,
910
+ 1
911
+ ]
912
+ ],
913
+ "flatten_loops": [
914
+ true
915
+ ],
916
+ "l2_groupings": [
917
+ 1
918
+ ],
919
+ "range_unroll_factors": [
920
+ 0
921
+ ],
922
+ "range_warp_specializes": [],
923
+ "range_num_stages": [
924
+ 0
925
+ ],
926
+ "range_multi_buffers": [
927
+ null
928
+ ],
929
+ "range_flattens": [
930
+ null
931
+ ],
932
+ "load_eviction_policies": [
933
+ "",
934
+ "",
935
+ ""
936
+ ],
937
+ "num_warps": 2,
938
+ "num_stages": 1,
939
+ "indexing": [
940
+ "pointer",
941
+ "pointer",
942
+ "pointer",
943
+ "pointer"
944
+ ],
945
+ "pid_type": "flat"
946
+ },
947
+ "intermediate_2880_numtokens_4": {
948
+ "block_sizes": [
949
+ 4,
950
+ 64
951
+ ],
952
+ "loop_orders": [
953
+ [
954
+ 0,
955
+ 1
956
+ ]
957
+ ],
958
+ "flatten_loops": [
959
+ true
960
+ ],
961
+ "l2_groupings": [
962
+ 64
963
+ ],
964
+ "range_unroll_factors": [
965
+ 0
966
+ ],
967
+ "range_warp_specializes": [],
968
+ "range_num_stages": [
969
+ 0
970
+ ],
971
+ "range_multi_buffers": [
972
+ null
973
+ ],
974
+ "range_flattens": [
975
+ null
976
+ ],
977
+ "load_eviction_policies": [
978
+ "",
979
+ "last",
980
+ "first"
981
+ ],
982
+ "num_warps": 8,
983
+ "num_stages": 7,
984
+ "indexing": [
985
+ "pointer",
986
+ "tensor_descriptor",
987
+ "tensor_descriptor",
988
+ "pointer"
989
+ ],
990
+ "pid_type": "xyz"
991
+ },
992
+ "intermediate_4096_numtokens_4": {
993
+ "block_sizes": [
994
+ 4,
995
+ 64
996
+ ],
997
+ "loop_orders": [
998
+ [
999
+ 0,
1000
+ 1
1001
+ ]
1002
+ ],
1003
+ "flatten_loops": [
1004
+ true
1005
+ ],
1006
+ "l2_groupings": [
1007
+ 1
1008
+ ],
1009
+ "range_unroll_factors": [
1010
+ 0
1011
+ ],
1012
+ "range_warp_specializes": [],
1013
+ "range_num_stages": [
1014
+ 0
1015
+ ],
1016
+ "range_multi_buffers": [
1017
+ null
1018
+ ],
1019
+ "range_flattens": [
1020
+ null
1021
+ ],
1022
+ "load_eviction_policies": [
1023
+ "",
1024
+ "",
1025
+ ""
1026
+ ],
1027
+ "num_warps": 2,
1028
+ "num_stages": 1,
1029
+ "indexing": [
1030
+ "pointer",
1031
+ "pointer",
1032
+ "pointer",
1033
+ "pointer"
1034
+ ],
1035
+ "pid_type": "flat"
1036
+ },
1037
+ "intermediate_8192_numtokens_4": {
1038
+ "block_sizes": [
1039
+ 1,
1040
+ 2048
1041
+ ],
1042
+ "loop_orders": [
1043
+ [
1044
+ 0,
1045
+ 1
1046
+ ]
1047
+ ],
1048
+ "flatten_loops": [
1049
+ false
1050
+ ],
1051
+ "l2_groupings": [
1052
+ 1
1053
+ ],
1054
+ "range_unroll_factors": [
1055
+ 0
1056
+ ],
1057
+ "range_warp_specializes": [],
1058
+ "range_num_stages": [
1059
+ 0
1060
+ ],
1061
+ "range_multi_buffers": [
1062
+ null
1063
+ ],
1064
+ "range_flattens": [
1065
+ null
1066
+ ],
1067
+ "load_eviction_policies": [
1068
+ "",
1069
+ "",
1070
+ ""
1071
+ ],
1072
+ "num_warps": 8,
1073
+ "num_stages": 8,
1074
+ "indexing": [
1075
+ "pointer",
1076
+ "tensor_descriptor",
1077
+ "pointer",
1078
+ "pointer"
1079
+ ],
1080
+ "pid_type": "flat"
1081
+ },
1082
+ "intermediate_11008_numtokens_4": {
1083
+ "block_sizes": [
1084
+ 4,
1085
+ 2048
1086
+ ],
1087
+ "loop_orders": [
1088
+ [
1089
+ 1,
1090
+ 0
1091
+ ]
1092
+ ],
1093
+ "flatten_loops": [
1094
+ true
1095
+ ],
1096
+ "l2_groupings": [
1097
+ 4
1098
+ ],
1099
+ "range_unroll_factors": [
1100
+ 0
1101
+ ],
1102
+ "range_warp_specializes": [],
1103
+ "range_num_stages": [
1104
+ 0
1105
+ ],
1106
+ "range_multi_buffers": [
1107
+ null
1108
+ ],
1109
+ "range_flattens": [
1110
+ null
1111
+ ],
1112
+ "load_eviction_policies": [
1113
+ "first",
1114
+ "last",
1115
+ "last"
1116
+ ],
1117
+ "num_warps": 8,
1118
+ "num_stages": 6,
1119
+ "indexing": [
1120
+ "pointer",
1121
+ "tensor_descriptor",
1122
+ "tensor_descriptor",
1123
+ "tensor_descriptor"
1124
+ ],
1125
+ "pid_type": "xyz"
1126
+ },
1127
+ "intermediate_14336_numtokens_4": {
1128
+ "block_sizes": [
1129
+ 4,
1130
+ 256
1131
+ ],
1132
+ "loop_orders": [
1133
+ [
1134
+ 0,
1135
+ 1
1136
+ ]
1137
+ ],
1138
+ "flatten_loops": [
1139
+ true
1140
+ ],
1141
+ "l2_groupings": [
1142
+ 64
1143
+ ],
1144
+ "range_unroll_factors": [
1145
+ 0
1146
+ ],
1147
+ "range_warp_specializes": [],
1148
+ "range_num_stages": [
1149
+ 0
1150
+ ],
1151
+ "range_multi_buffers": [
1152
+ null
1153
+ ],
1154
+ "range_flattens": [
1155
+ null
1156
+ ],
1157
+ "load_eviction_policies": [
1158
+ "",
1159
+ "",
1160
+ ""
1161
+ ],
1162
+ "num_warps": 2,
1163
+ "num_stages": 2,
1164
+ "indexing": [
1165
+ "pointer",
1166
+ "pointer",
1167
+ "pointer",
1168
+ "pointer"
1169
+ ],
1170
+ "pid_type": "xyz"
1171
+ },
1172
+ "intermediate_2048_numtokens_8": {
1173
+ "block_sizes": [
1174
+ 8,
1175
+ 32
1176
+ ],
1177
+ "loop_orders": [
1178
+ [
1179
+ 0,
1180
+ 1
1181
+ ]
1182
+ ],
1183
+ "flatten_loops": [
1184
+ true
1185
+ ],
1186
+ "l2_groupings": [
1187
+ 1
1188
+ ],
1189
+ "range_unroll_factors": [
1190
+ 0
1191
+ ],
1192
+ "range_warp_specializes": [],
1193
+ "range_num_stages": [
1194
+ 0
1195
+ ],
1196
+ "range_multi_buffers": [
1197
+ null
1198
+ ],
1199
+ "range_flattens": [
1200
+ null
1201
+ ],
1202
+ "load_eviction_policies": [
1203
+ "",
1204
+ "",
1205
+ ""
1206
+ ],
1207
+ "num_warps": 2,
1208
+ "num_stages": 1,
1209
+ "indexing": [
1210
+ "pointer",
1211
+ "pointer",
1212
+ "pointer",
1213
+ "pointer"
1214
+ ],
1215
+ "pid_type": "flat"
1216
+ },
1217
+ "intermediate_2880_numtokens_8": {
1218
+ "block_sizes": [
1219
+ 4,
1220
+ 2048
1221
+ ],
1222
+ "loop_orders": [
1223
+ [
1224
+ 0,
1225
+ 1
1226
+ ]
1227
+ ],
1228
+ "flatten_loops": [
1229
+ true
1230
+ ],
1231
+ "l2_groupings": [
1232
+ 4
1233
+ ],
1234
+ "range_unroll_factors": [
1235
+ 0
1236
+ ],
1237
+ "range_warp_specializes": [],
1238
+ "range_num_stages": [
1239
+ 0
1240
+ ],
1241
+ "range_multi_buffers": [
1242
+ null
1243
+ ],
1244
+ "range_flattens": [
1245
+ null
1246
+ ],
1247
+ "load_eviction_policies": [
1248
+ "last",
1249
+ "last",
1250
+ ""
1251
+ ],
1252
+ "num_warps": 16,
1253
+ "num_stages": 5,
1254
+ "indexing": [
1255
+ "pointer",
1256
+ "tensor_descriptor",
1257
+ "tensor_descriptor",
1258
+ "pointer"
1259
+ ],
1260
+ "pid_type": "xyz"
1261
+ },
1262
+ "intermediate_4096_numtokens_8": {
1263
+ "block_sizes": [
1264
+ 8,
1265
+ 32
1266
+ ],
1267
+ "loop_orders": [
1268
+ [
1269
+ 0,
1270
+ 1
1271
+ ]
1272
+ ],
1273
+ "flatten_loops": [
1274
+ true
1275
+ ],
1276
+ "l2_groupings": [
1277
+ 1
1278
+ ],
1279
+ "range_unroll_factors": [
1280
+ 0
1281
+ ],
1282
+ "range_warp_specializes": [],
1283
+ "range_num_stages": [
1284
+ 0
1285
+ ],
1286
+ "range_multi_buffers": [
1287
+ null
1288
+ ],
1289
+ "range_flattens": [
1290
+ null
1291
+ ],
1292
+ "load_eviction_policies": [
1293
+ "",
1294
+ "",
1295
+ ""
1296
+ ],
1297
+ "num_warps": 2,
1298
+ "num_stages": 1,
1299
+ "indexing": [
1300
+ "pointer",
1301
+ "pointer",
1302
+ "pointer",
1303
+ "pointer"
1304
+ ],
1305
+ "pid_type": "flat"
1306
+ },
1307
+ "intermediate_8192_numtokens_8": {
1308
+ "block_sizes": [
1309
+ 2,
1310
+ 1024
1311
+ ],
1312
+ "loop_orders": [
1313
+ [
1314
+ 0,
1315
+ 1
1316
+ ]
1317
+ ],
1318
+ "flatten_loops": [
1319
+ true
1320
+ ],
1321
+ "l2_groupings": [
1322
+ 16
1323
+ ],
1324
+ "range_unroll_factors": [
1325
+ 0
1326
+ ],
1327
+ "range_warp_specializes": [],
1328
+ "range_num_stages": [
1329
+ 0
1330
+ ],
1331
+ "range_multi_buffers": [
1332
+ null
1333
+ ],
1334
+ "range_flattens": [
1335
+ null
1336
+ ],
1337
+ "load_eviction_policies": [
1338
+ "first",
1339
+ "last",
1340
+ ""
1341
+ ],
1342
+ "num_warps": 1,
1343
+ "num_stages": 8,
1344
+ "indexing": [
1345
+ "tensor_descriptor",
1346
+ "pointer",
1347
+ "tensor_descriptor",
1348
+ "pointer"
1349
+ ],
1350
+ "pid_type": "flat"
1351
+ },
1352
+ "intermediate_11008_numtokens_8": {
1353
+ "block_sizes": [
1354
+ 1,
1355
+ 1024
1356
+ ],
1357
+ "loop_orders": [
1358
+ [
1359
+ 1,
1360
+ 0
1361
+ ]
1362
+ ],
1363
+ "flatten_loops": [
1364
+ true
1365
+ ],
1366
+ "l2_groupings": [
1367
+ 4
1368
+ ],
1369
+ "range_unroll_factors": [
1370
+ 0
1371
+ ],
1372
+ "range_warp_specializes": [],
1373
+ "range_num_stages": [
1374
+ 0
1375
+ ],
1376
+ "range_multi_buffers": [
1377
+ null
1378
+ ],
1379
+ "range_flattens": [
1380
+ null
1381
+ ],
1382
+ "load_eviction_policies": [
1383
+ "last",
1384
+ "",
1385
+ "first"
1386
+ ],
1387
+ "num_warps": 2,
1388
+ "num_stages": 5,
1389
+ "indexing": [
1390
+ "tensor_descriptor",
1391
+ "tensor_descriptor",
1392
+ "tensor_descriptor",
1393
+ "tensor_descriptor"
1394
+ ],
1395
+ "pid_type": "flat"
1396
+ },
1397
+ "intermediate_14336_numtokens_8": {
1398
+ "block_sizes": [
1399
+ 8,
1400
+ 128
1401
+ ],
1402
+ "loop_orders": [
1403
+ [
1404
+ 0,
1405
+ 1
1406
+ ]
1407
+ ],
1408
+ "flatten_loops": [
1409
+ true
1410
+ ],
1411
+ "l2_groupings": [
1412
+ 1
1413
+ ],
1414
+ "range_unroll_factors": [
1415
+ 0
1416
+ ],
1417
+ "range_warp_specializes": [],
1418
+ "range_num_stages": [
1419
+ 0
1420
+ ],
1421
+ "range_multi_buffers": [
1422
+ null
1423
+ ],
1424
+ "range_flattens": [
1425
+ null
1426
+ ],
1427
+ "load_eviction_policies": [
1428
+ "",
1429
+ "",
1430
+ ""
1431
+ ],
1432
+ "num_warps": 4,
1433
+ "num_stages": 1,
1434
+ "indexing": [
1435
+ "pointer",
1436
+ "pointer",
1437
+ "pointer",
1438
+ "pointer"
1439
+ ],
1440
+ "pid_type": "flat"
1441
+ },
1442
+ "intermediate_2048_numtokens_16": {
1443
+ "block_sizes": [
1444
+ 8,
1445
+ 512
1446
+ ],
1447
+ "loop_orders": [
1448
+ [
1449
+ 0,
1450
+ 1
1451
+ ]
1452
+ ],
1453
+ "flatten_loops": [
1454
+ true
1455
+ ],
1456
+ "l2_groupings": [
1457
+ 64
1458
+ ],
1459
+ "range_unroll_factors": [
1460
+ 0
1461
+ ],
1462
+ "range_warp_specializes": [],
1463
+ "range_num_stages": [
1464
+ 0
1465
+ ],
1466
+ "range_multi_buffers": [
1467
+ null
1468
+ ],
1469
+ "range_flattens": [
1470
+ null
1471
+ ],
1472
+ "load_eviction_policies": [
1473
+ "last",
1474
+ "last",
1475
+ "first"
1476
+ ],
1477
+ "num_warps": 16,
1478
+ "num_stages": 2,
1479
+ "indexing": [
1480
+ "pointer",
1481
+ "pointer",
1482
+ "tensor_descriptor",
1483
+ "pointer"
1484
+ ],
1485
+ "pid_type": "xyz"
1486
+ },
1487
+ "intermediate_2880_numtokens_16": {
1488
+ "block_sizes": [
1489
+ 2,
1490
+ 256
1491
+ ],
1492
+ "loop_orders": [
1493
+ [
1494
+ 0,
1495
+ 1
1496
+ ]
1497
+ ],
1498
+ "flatten_loops": [
1499
+ true
1500
+ ],
1501
+ "l2_groupings": [
1502
+ 1
1503
+ ],
1504
+ "range_unroll_factors": [
1505
+ 0
1506
+ ],
1507
+ "range_warp_specializes": [],
1508
+ "range_num_stages": [
1509
+ 0
1510
+ ],
1511
+ "range_multi_buffers": [
1512
+ null
1513
+ ],
1514
+ "range_flattens": [
1515
+ null
1516
+ ],
1517
+ "load_eviction_policies": [
1518
+ "",
1519
+ "",
1520
+ ""
1521
+ ],
1522
+ "num_warps": 4,
1523
+ "num_stages": 1,
1524
+ "indexing": [
1525
+ "pointer",
1526
+ "pointer",
1527
+ "pointer",
1528
+ "pointer"
1529
+ ],
1530
+ "pid_type": "flat"
1531
+ },
1532
+ "intermediate_4096_numtokens_16": {
1533
+ "block_sizes": [
1534
+ 16,
1535
+ 256
1536
+ ],
1537
+ "loop_orders": [
1538
+ [
1539
+ 0,
1540
+ 1
1541
+ ]
1542
+ ],
1543
+ "flatten_loops": [
1544
+ true
1545
+ ],
1546
+ "l2_groupings": [
1547
+ 16
1548
+ ],
1549
+ "range_unroll_factors": [
1550
+ 0
1551
+ ],
1552
+ "range_warp_specializes": [],
1553
+ "range_num_stages": [
1554
+ 0
1555
+ ],
1556
+ "range_multi_buffers": [
1557
+ null
1558
+ ],
1559
+ "range_flattens": [
1560
+ null
1561
+ ],
1562
+ "load_eviction_policies": [
1563
+ "last",
1564
+ "",
1565
+ "last"
1566
+ ],
1567
+ "num_warps": 32,
1568
+ "num_stages": 1,
1569
+ "indexing": [
1570
+ "pointer",
1571
+ "pointer",
1572
+ "pointer",
1573
+ "pointer"
1574
+ ],
1575
+ "pid_type": "flat"
1576
+ },
1577
+ "intermediate_8192_numtokens_16": {
1578
+ "block_sizes": [
1579
+ 16,
1580
+ 64
1581
+ ],
1582
+ "loop_orders": [
1583
+ [
1584
+ 0,
1585
+ 1
1586
+ ]
1587
+ ],
1588
+ "flatten_loops": [
1589
+ true
1590
+ ],
1591
+ "l2_groupings": [
1592
+ 1
1593
+ ],
1594
+ "range_unroll_factors": [
1595
+ 0
1596
+ ],
1597
+ "range_warp_specializes": [],
1598
+ "range_num_stages": [
1599
+ 0
1600
+ ],
1601
+ "range_multi_buffers": [
1602
+ null
1603
+ ],
1604
+ "range_flattens": [
1605
+ null
1606
+ ],
1607
+ "load_eviction_policies": [
1608
+ "",
1609
+ "",
1610
+ ""
1611
+ ],
1612
+ "num_warps": 4,
1613
+ "num_stages": 1,
1614
+ "indexing": [
1615
+ "pointer",
1616
+ "pointer",
1617
+ "pointer",
1618
+ "pointer"
1619
+ ],
1620
+ "pid_type": "flat"
1621
+ },
1622
+ "intermediate_11008_numtokens_16": {
1623
+ "block_sizes": [
1624
+ 1,
1625
+ 2048
1626
+ ],
1627
+ "loop_orders": [
1628
+ [
1629
+ 0,
1630
+ 1
1631
+ ]
1632
+ ],
1633
+ "flatten_loops": [
1634
+ true
1635
+ ],
1636
+ "l2_groupings": [
1637
+ 1
1638
+ ],
1639
+ "range_unroll_factors": [
1640
+ 0
1641
+ ],
1642
+ "range_warp_specializes": [],
1643
+ "range_num_stages": [
1644
+ 0
1645
+ ],
1646
+ "range_multi_buffers": [
1647
+ null
1648
+ ],
1649
+ "range_flattens": [
1650
+ null
1651
+ ],
1652
+ "load_eviction_policies": [
1653
+ "",
1654
+ "",
1655
+ ""
1656
+ ],
1657
+ "num_warps": 8,
1658
+ "num_stages": 1,
1659
+ "indexing": [
1660
+ "pointer",
1661
+ "pointer",
1662
+ "pointer",
1663
+ "pointer"
1664
+ ],
1665
+ "pid_type": "flat"
1666
+ },
1667
+ "intermediate_14336_numtokens_16": {
1668
+ "block_sizes": [
1669
+ 2,
1670
+ 256
1671
+ ],
1672
+ "loop_orders": [
1673
+ [
1674
+ 1,
1675
+ 0
1676
+ ]
1677
+ ],
1678
+ "flatten_loops": [
1679
+ true
1680
+ ],
1681
+ "l2_groupings": [
1682
+ 64
1683
+ ],
1684
+ "range_unroll_factors": [
1685
+ 0
1686
+ ],
1687
+ "range_warp_specializes": [],
1688
+ "range_num_stages": [
1689
+ 0
1690
+ ],
1691
+ "range_multi_buffers": [
1692
+ null
1693
+ ],
1694
+ "range_flattens": [
1695
+ null
1696
+ ],
1697
+ "load_eviction_policies": [
1698
+ "last",
1699
+ "last",
1700
+ "last"
1701
+ ],
1702
+ "num_warps": 16,
1703
+ "num_stages": 7,
1704
+ "indexing": [
1705
+ "tensor_descriptor",
1706
+ "pointer",
1707
+ "tensor_descriptor",
1708
+ "tensor_descriptor"
1709
+ ],
1710
+ "pid_type": "flat"
1711
+ },
1712
+ "intermediate_2048_numtokens_24": {
1713
+ "block_sizes": [
1714
+ 1,
1715
+ 1024
1716
+ ],
1717
+ "loop_orders": [
1718
+ [
1719
+ 0,
1720
+ 1
1721
+ ]
1722
+ ],
1723
+ "flatten_loops": [
1724
+ true
1725
+ ],
1726
+ "l2_groupings": [
1727
+ 64
1728
+ ],
1729
+ "range_unroll_factors": [
1730
+ 0
1731
+ ],
1732
+ "range_warp_specializes": [],
1733
+ "range_num_stages": [
1734
+ 0
1735
+ ],
1736
+ "range_multi_buffers": [
1737
+ null
1738
+ ],
1739
+ "range_flattens": [
1740
+ null
1741
+ ],
1742
+ "load_eviction_policies": [
1743
+ "",
1744
+ "",
1745
+ "first"
1746
+ ],
1747
+ "num_warps": 4,
1748
+ "num_stages": 8,
1749
+ "indexing": [
1750
+ "tensor_descriptor",
1751
+ "pointer",
1752
+ "pointer",
1753
+ "tensor_descriptor"
1754
+ ],
1755
+ "pid_type": "flat"
1756
+ },
1757
+ "intermediate_2880_numtokens_24": {
1758
+ "block_sizes": [
1759
+ 4,
1760
+ 1024
1761
+ ],
1762
+ "loop_orders": [
1763
+ [
1764
+ 0,
1765
+ 1
1766
+ ]
1767
+ ],
1768
+ "flatten_loops": [
1769
+ true
1770
+ ],
1771
+ "l2_groupings": [
1772
+ 16
1773
+ ],
1774
+ "range_unroll_factors": [
1775
+ 0
1776
+ ],
1777
+ "range_warp_specializes": [],
1778
+ "range_num_stages": [
1779
+ 0
1780
+ ],
1781
+ "range_multi_buffers": [
1782
+ null
1783
+ ],
1784
+ "range_flattens": [
1785
+ null
1786
+ ],
1787
+ "load_eviction_policies": [
1788
+ "first",
1789
+ "",
1790
+ "first"
1791
+ ],
1792
+ "num_warps": 16,
1793
+ "num_stages": 3,
1794
+ "indexing": [
1795
+ "tensor_descriptor",
1796
+ "pointer",
1797
+ "tensor_descriptor",
1798
+ "pointer"
1799
+ ],
1800
+ "pid_type": "flat"
1801
+ },
1802
+ "intermediate_4096_numtokens_24": {
1803
+ "block_sizes": [
1804
+ 16,
1805
+ 64
1806
+ ],
1807
+ "loop_orders": [
1808
+ [
1809
+ 0,
1810
+ 1
1811
+ ]
1812
+ ],
1813
+ "flatten_loops": [
1814
+ true
1815
+ ],
1816
+ "l2_groupings": [
1817
+ 1
1818
+ ],
1819
+ "range_unroll_factors": [
1820
+ 0
1821
+ ],
1822
+ "range_warp_specializes": [],
1823
+ "range_num_stages": [
1824
+ 0
1825
+ ],
1826
+ "range_multi_buffers": [
1827
+ null
1828
+ ],
1829
+ "range_flattens": [
1830
+ null
1831
+ ],
1832
+ "load_eviction_policies": [
1833
+ "",
1834
+ "",
1835
+ ""
1836
+ ],
1837
+ "num_warps": 4,
1838
+ "num_stages": 1,
1839
+ "indexing": [
1840
+ "pointer",
1841
+ "pointer",
1842
+ "pointer",
1843
+ "pointer"
1844
+ ],
1845
+ "pid_type": "flat"
1846
+ },
1847
+ "intermediate_8192_numtokens_24": {
1848
+ "block_sizes": [
1849
+ 1,
1850
+ 2048
1851
+ ],
1852
+ "loop_orders": [
1853
+ [
1854
+ 0,
1855
+ 1
1856
+ ]
1857
+ ],
1858
+ "flatten_loops": [
1859
+ true
1860
+ ],
1861
+ "l2_groupings": [
1862
+ 32
1863
+ ],
1864
+ "range_unroll_factors": [
1865
+ 0
1866
+ ],
1867
+ "range_warp_specializes": [],
1868
+ "range_num_stages": [
1869
+ 0
1870
+ ],
1871
+ "range_multi_buffers": [
1872
+ null
1873
+ ],
1874
+ "range_flattens": [
1875
+ null
1876
+ ],
1877
+ "load_eviction_policies": [
1878
+ "",
1879
+ "last",
1880
+ "last"
1881
+ ],
1882
+ "num_warps": 32,
1883
+ "num_stages": 5,
1884
+ "indexing": [
1885
+ "pointer",
1886
+ "pointer",
1887
+ "pointer",
1888
+ "tensor_descriptor"
1889
+ ],
1890
+ "pid_type": "flat"
1891
+ },
1892
+ "intermediate_11008_numtokens_24": {
1893
+ "block_sizes": [
1894
+ 1,
1895
+ 1024
1896
+ ],
1897
+ "loop_orders": [
1898
+ [
1899
+ 1,
1900
+ 0
1901
+ ]
1902
+ ],
1903
+ "flatten_loops": [
1904
+ false
1905
+ ],
1906
+ "l2_groupings": [
1907
+ 64
1908
+ ],
1909
+ "range_unroll_factors": [
1910
+ 0
1911
+ ],
1912
+ "range_warp_specializes": [],
1913
+ "range_num_stages": [
1914
+ 0
1915
+ ],
1916
+ "range_multi_buffers": [
1917
+ null
1918
+ ],
1919
+ "range_flattens": [
1920
+ null
1921
+ ],
1922
+ "load_eviction_policies": [
1923
+ "last",
1924
+ "",
1925
+ ""
1926
+ ],
1927
+ "num_warps": 2,
1928
+ "num_stages": 2,
1929
+ "indexing": [
1930
+ "pointer",
1931
+ "pointer",
1932
+ "pointer",
1933
+ "pointer"
1934
+ ],
1935
+ "pid_type": "flat"
1936
+ },
1937
+ "intermediate_14336_numtokens_24": {
1938
+ "block_sizes": [
1939
+ 8,
1940
+ 512
1941
+ ],
1942
+ "loop_orders": [
1943
+ [
1944
+ 0,
1945
+ 1
1946
+ ]
1947
+ ],
1948
+ "flatten_loops": [
1949
+ true
1950
+ ],
1951
+ "l2_groupings": [
1952
+ 4
1953
+ ],
1954
+ "range_unroll_factors": [
1955
+ 0
1956
+ ],
1957
+ "range_warp_specializes": [],
1958
+ "range_num_stages": [
1959
+ 0
1960
+ ],
1961
+ "range_multi_buffers": [
1962
+ null
1963
+ ],
1964
+ "range_flattens": [
1965
+ null
1966
+ ],
1967
+ "load_eviction_policies": [
1968
+ "last",
1969
+ "first",
1970
+ "first"
1971
+ ],
1972
+ "num_warps": 32,
1973
+ "num_stages": 3,
1974
+ "indexing": [
1975
+ "pointer",
1976
+ "tensor_descriptor",
1977
+ "pointer",
1978
+ "tensor_descriptor"
1979
+ ],
1980
+ "pid_type": "flat"
1981
+ },
1982
+ "intermediate_2048_numtokens_32": {
1983
+ "block_sizes": [
1984
+ 32,
1985
+ 16
1986
+ ],
1987
+ "loop_orders": [
1988
+ [
1989
+ 0,
1990
+ 1
1991
+ ]
1992
+ ],
1993
+ "flatten_loops": [
1994
+ true
1995
+ ],
1996
+ "l2_groupings": [
1997
+ 1
1998
+ ],
1999
+ "range_unroll_factors": [
2000
+ 0
2001
+ ],
2002
+ "range_warp_specializes": [],
2003
+ "range_num_stages": [
2004
+ 0
2005
+ ],
2006
+ "range_multi_buffers": [
2007
+ null
2008
+ ],
2009
+ "range_flattens": [
2010
+ null
2011
+ ],
2012
+ "load_eviction_policies": [
2013
+ "",
2014
+ "",
2015
+ ""
2016
+ ],
2017
+ "num_warps": 4,
2018
+ "num_stages": 1,
2019
+ "indexing": [
2020
+ "pointer",
2021
+ "pointer",
2022
+ "pointer",
2023
+ "pointer"
2024
+ ],
2025
+ "pid_type": "flat"
2026
+ },
2027
+ "intermediate_2880_numtokens_32": {
2028
+ "block_sizes": [
2029
+ 4,
2030
+ 256
2031
+ ],
2032
+ "loop_orders": [
2033
+ [
2034
+ 0,
2035
+ 1
2036
+ ]
2037
+ ],
2038
+ "flatten_loops": [
2039
+ true
2040
+ ],
2041
+ "l2_groupings": [
2042
+ 1
2043
+ ],
2044
+ "range_unroll_factors": [
2045
+ 0
2046
+ ],
2047
+ "range_warp_specializes": [],
2048
+ "range_num_stages": [
2049
+ 0
2050
+ ],
2051
+ "range_multi_buffers": [
2052
+ null
2053
+ ],
2054
+ "range_flattens": [
2055
+ null
2056
+ ],
2057
+ "load_eviction_policies": [
2058
+ "",
2059
+ "",
2060
+ ""
2061
+ ],
2062
+ "num_warps": 4,
2063
+ "num_stages": 1,
2064
+ "indexing": [
2065
+ "pointer",
2066
+ "pointer",
2067
+ "pointer",
2068
+ "pointer"
2069
+ ],
2070
+ "pid_type": "flat"
2071
+ },
2072
+ "intermediate_4096_numtokens_32": {
2073
+ "block_sizes": [
2074
+ 4,
2075
+ 4096
2076
+ ],
2077
+ "loop_orders": [
2078
+ [
2079
+ 0,
2080
+ 1
2081
+ ]
2082
+ ],
2083
+ "flatten_loops": [
2084
+ true
2085
+ ],
2086
+ "l2_groupings": [
2087
+ 16
2088
+ ],
2089
+ "range_unroll_factors": [
2090
+ 0
2091
+ ],
2092
+ "range_warp_specializes": [],
2093
+ "range_num_stages": [
2094
+ 0
2095
+ ],
2096
+ "range_multi_buffers": [
2097
+ null
2098
+ ],
2099
+ "range_flattens": [
2100
+ null
2101
+ ],
2102
+ "load_eviction_policies": [
2103
+ "first",
2104
+ "last",
2105
+ ""
2106
+ ],
2107
+ "num_warps": 32,
2108
+ "num_stages": 5,
2109
+ "indexing": [
2110
+ "pointer",
2111
+ "tensor_descriptor",
2112
+ "pointer",
2113
+ "pointer"
2114
+ ],
2115
+ "pid_type": "flat"
2116
+ },
2117
+ "intermediate_8192_numtokens_32": {
2118
+ "block_sizes": [
2119
+ 4,
2120
+ 1024
2121
+ ],
2122
+ "loop_orders": [
2123
+ [
2124
+ 0,
2125
+ 1
2126
+ ]
2127
+ ],
2128
+ "flatten_loops": [
2129
+ true
2130
+ ],
2131
+ "l2_groupings": [
2132
+ 2
2133
+ ],
2134
+ "range_unroll_factors": [
2135
+ 0
2136
+ ],
2137
+ "range_warp_specializes": [],
2138
+ "range_num_stages": [
2139
+ 0
2140
+ ],
2141
+ "range_multi_buffers": [
2142
+ null
2143
+ ],
2144
+ "range_flattens": [
2145
+ null
2146
+ ],
2147
+ "load_eviction_policies": [
2148
+ "",
2149
+ "last",
2150
+ ""
2151
+ ],
2152
+ "num_warps": 2,
2153
+ "num_stages": 3,
2154
+ "indexing": [
2155
+ "tensor_descriptor",
2156
+ "tensor_descriptor",
2157
+ "pointer",
2158
+ "tensor_descriptor"
2159
+ ],
2160
+ "pid_type": "flat"
2161
+ },
2162
+ "intermediate_11008_numtokens_32": {
2163
+ "block_sizes": [
2164
+ 2,
2165
+ 2048
2166
+ ],
2167
+ "loop_orders": [
2168
+ [
2169
+ 0,
2170
+ 1
2171
+ ]
2172
+ ],
2173
+ "flatten_loops": [
2174
+ true
2175
+ ],
2176
+ "l2_groupings": [
2177
+ 1
2178
+ ],
2179
+ "range_unroll_factors": [
2180
+ 0
2181
+ ],
2182
+ "range_warp_specializes": [],
2183
+ "range_num_stages": [
2184
+ 0
2185
+ ],
2186
+ "range_multi_buffers": [
2187
+ null
2188
+ ],
2189
+ "range_flattens": [
2190
+ null
2191
+ ],
2192
+ "load_eviction_policies": [
2193
+ "",
2194
+ "",
2195
+ ""
2196
+ ],
2197
+ "num_warps": 8,
2198
+ "num_stages": 1,
2199
+ "indexing": [
2200
+ "pointer",
2201
+ "pointer",
2202
+ "pointer",
2203
+ "pointer"
2204
+ ],
2205
+ "pid_type": "flat"
2206
+ },
2207
+ "intermediate_14336_numtokens_32": {
2208
+ "block_sizes": [
2209
+ 1,
2210
+ 512
2211
+ ],
2212
+ "loop_orders": [
2213
+ [
2214
+ 0,
2215
+ 1
2216
+ ]
2217
+ ],
2218
+ "flatten_loops": [
2219
+ true
2220
+ ],
2221
+ "l2_groupings": [
2222
+ 64
2223
+ ],
2224
+ "range_unroll_factors": [
2225
+ 0
2226
+ ],
2227
+ "range_warp_specializes": [],
2228
+ "range_num_stages": [
2229
+ 0
2230
+ ],
2231
+ "range_multi_buffers": [
2232
+ null
2233
+ ],
2234
+ "range_flattens": [
2235
+ null
2236
+ ],
2237
+ "load_eviction_policies": [
2238
+ "first",
2239
+ "",
2240
+ ""
2241
+ ],
2242
+ "num_warps": 4,
2243
+ "num_stages": 3,
2244
+ "indexing": [
2245
+ "pointer",
2246
+ "tensor_descriptor",
2247
+ "pointer",
2248
+ "pointer"
2249
+ ],
2250
+ "pid_type": "flat"
2251
+ },
2252
+ "intermediate_2048_numtokens_40": {
2253
+ "block_sizes": [
2254
+ 32,
2255
+ 32
2256
+ ],
2257
+ "loop_orders": [
2258
+ [
2259
+ 0,
2260
+ 1
2261
+ ]
2262
+ ],
2263
+ "flatten_loops": [
2264
+ true
2265
+ ],
2266
+ "l2_groupings": [
2267
+ 1
2268
+ ],
2269
+ "range_unroll_factors": [
2270
+ 0
2271
+ ],
2272
+ "range_warp_specializes": [],
2273
+ "range_num_stages": [
2274
+ 0
2275
+ ],
2276
+ "range_multi_buffers": [
2277
+ null
2278
+ ],
2279
+ "range_flattens": [
2280
+ null
2281
+ ],
2282
+ "load_eviction_policies": [
2283
+ "",
2284
+ "",
2285
+ ""
2286
+ ],
2287
+ "num_warps": 4,
2288
+ "num_stages": 1,
2289
+ "indexing": [
2290
+ "pointer",
2291
+ "pointer",
2292
+ "pointer",
2293
+ "pointer"
2294
+ ],
2295
+ "pid_type": "flat"
2296
+ },
2297
+ "intermediate_2880_numtokens_40": {
2298
+ "block_sizes": [
2299
+ 1,
2300
+ 4096
2301
+ ],
2302
+ "loop_orders": [
2303
+ [
2304
+ 1,
2305
+ 0
2306
+ ]
2307
+ ],
2308
+ "flatten_loops": [
2309
+ false
2310
+ ],
2311
+ "l2_groupings": [
2312
+ 32
2313
+ ],
2314
+ "range_unroll_factors": [
2315
+ 0
2316
+ ],
2317
+ "range_warp_specializes": [],
2318
+ "range_num_stages": [
2319
+ 0
2320
+ ],
2321
+ "range_multi_buffers": [
2322
+ null
2323
+ ],
2324
+ "range_flattens": [
2325
+ null
2326
+ ],
2327
+ "load_eviction_policies": [
2328
+ "",
2329
+ "last",
2330
+ "last"
2331
+ ],
2332
+ "num_warps": 8,
2333
+ "num_stages": 4,
2334
+ "indexing": [
2335
+ "tensor_descriptor",
2336
+ "pointer",
2337
+ "pointer",
2338
+ "pointer"
2339
+ ],
2340
+ "pid_type": "flat"
2341
+ },
2342
+ "intermediate_4096_numtokens_40": {
2343
+ "block_sizes": [
2344
+ 32,
2345
+ 64
2346
+ ],
2347
+ "loop_orders": [
2348
+ [
2349
+ 0,
2350
+ 1
2351
+ ]
2352
+ ],
2353
+ "flatten_loops": [
2354
+ true
2355
+ ],
2356
+ "l2_groupings": [
2357
+ 1
2358
+ ],
2359
+ "range_unroll_factors": [
2360
+ 0
2361
+ ],
2362
+ "range_warp_specializes": [],
2363
+ "range_num_stages": [
2364
+ 0
2365
+ ],
2366
+ "range_multi_buffers": [
2367
+ null
2368
+ ],
2369
+ "range_flattens": [
2370
+ null
2371
+ ],
2372
+ "load_eviction_policies": [
2373
+ "",
2374
+ "",
2375
+ ""
2376
+ ],
2377
+ "num_warps": 8,
2378
+ "num_stages": 1,
2379
+ "indexing": [
2380
+ "pointer",
2381
+ "pointer",
2382
+ "pointer",
2383
+ "pointer"
2384
+ ],
2385
+ "pid_type": "flat"
2386
+ },
2387
+ "intermediate_8192_numtokens_40": {
2388
+ "block_sizes": [
2389
+ 2,
2390
+ 256
2391
+ ],
2392
+ "loop_orders": [
2393
+ [
2394
+ 0,
2395
+ 1
2396
+ ]
2397
+ ],
2398
+ "flatten_loops": [
2399
+ true
2400
+ ],
2401
+ "l2_groupings": [
2402
+ 16
2403
+ ],
2404
+ "range_unroll_factors": [
2405
+ 0
2406
+ ],
2407
+ "range_warp_specializes": [],
2408
+ "range_num_stages": [
2409
+ 0
2410
+ ],
2411
+ "range_multi_buffers": [
2412
+ null
2413
+ ],
2414
+ "range_flattens": [
2415
+ null
2416
+ ],
2417
+ "load_eviction_policies": [
2418
+ "",
2419
+ "last",
2420
+ "last"
2421
+ ],
2422
+ "num_warps": 2,
2423
+ "num_stages": 2,
2424
+ "indexing": [
2425
+ "pointer",
2426
+ "pointer",
2427
+ "pointer",
2428
+ "tensor_descriptor"
2429
+ ],
2430
+ "pid_type": "flat"
2431
+ },
2432
+ "intermediate_11008_numtokens_40": {
2433
+ "block_sizes": [
2434
+ 16,
2435
+ 256
2436
+ ],
2437
+ "loop_orders": [
2438
+ [
2439
+ 0,
2440
+ 1
2441
+ ]
2442
+ ],
2443
+ "flatten_loops": [
2444
+ true
2445
+ ],
2446
+ "l2_groupings": [
2447
+ 1
2448
+ ],
2449
+ "range_unroll_factors": [
2450
+ 0
2451
+ ],
2452
+ "range_warp_specializes": [],
2453
+ "range_num_stages": [
2454
+ 0
2455
+ ],
2456
+ "range_multi_buffers": [
2457
+ null
2458
+ ],
2459
+ "range_flattens": [
2460
+ null
2461
+ ],
2462
+ "load_eviction_policies": [
2463
+ "",
2464
+ "",
2465
+ ""
2466
+ ],
2467
+ "num_warps": 8,
2468
+ "num_stages": 1,
2469
+ "indexing": [
2470
+ "pointer",
2471
+ "pointer",
2472
+ "pointer",
2473
+ "pointer"
2474
+ ],
2475
+ "pid_type": "flat"
2476
+ },
2477
+ "intermediate_14336_numtokens_40": {
2478
+ "block_sizes": [
2479
+ 4,
2480
+ 1024
2481
+ ],
2482
+ "loop_orders": [
2483
+ [
2484
+ 0,
2485
+ 1
2486
+ ]
2487
+ ],
2488
+ "flatten_loops": [
2489
+ true
2490
+ ],
2491
+ "l2_groupings": [
2492
+ 1
2493
+ ],
2494
+ "range_unroll_factors": [
2495
+ 0
2496
+ ],
2497
+ "range_warp_specializes": [],
2498
+ "range_num_stages": [
2499
+ 1
2500
+ ],
2501
+ "range_multi_buffers": [
2502
+ null
2503
+ ],
2504
+ "range_flattens": [
2505
+ null
2506
+ ],
2507
+ "load_eviction_policies": [
2508
+ "first",
2509
+ "",
2510
+ "last"
2511
+ ],
2512
+ "num_warps": 16,
2513
+ "num_stages": 5,
2514
+ "indexing": [
2515
+ "tensor_descriptor",
2516
+ "tensor_descriptor",
2517
+ "tensor_descriptor",
2518
+ "tensor_descriptor"
2519
+ ],
2520
+ "pid_type": "persistent_interleaved",
2521
+ "num_sm_multiplier": 32,
2522
+ "maxnreg": 32
2523
+ },
2524
+ "intermediate_2048_numtokens_48": {
2525
+ "block_sizes": [
2526
+ 32,
2527
+ 32
2528
+ ],
2529
+ "loop_orders": [
2530
+ [
2531
+ 0,
2532
+ 1
2533
+ ]
2534
+ ],
2535
+ "flatten_loops": [
2536
+ true
2537
+ ],
2538
+ "l2_groupings": [
2539
+ 1
2540
+ ],
2541
+ "range_unroll_factors": [
2542
+ 0
2543
+ ],
2544
+ "range_warp_specializes": [],
2545
+ "range_num_stages": [
2546
+ 0
2547
+ ],
2548
+ "range_multi_buffers": [
2549
+ null
2550
+ ],
2551
+ "range_flattens": [
2552
+ null
2553
+ ],
2554
+ "load_eviction_policies": [
2555
+ "",
2556
+ "",
2557
+ ""
2558
+ ],
2559
+ "num_warps": 4,
2560
+ "num_stages": 1,
2561
+ "indexing": [
2562
+ "pointer",
2563
+ "pointer",
2564
+ "pointer",
2565
+ "pointer"
2566
+ ],
2567
+ "pid_type": "flat"
2568
+ },
2569
+ "intermediate_2880_numtokens_48": {
2570
+ "block_sizes": [
2571
+ 16,
2572
+ 64
2573
+ ],
2574
+ "loop_orders": [
2575
+ [
2576
+ 0,
2577
+ 1
2578
+ ]
2579
+ ],
2580
+ "flatten_loops": [
2581
+ true
2582
+ ],
2583
+ "l2_groupings": [
2584
+ 1
2585
+ ],
2586
+ "range_unroll_factors": [
2587
+ 0
2588
+ ],
2589
+ "range_warp_specializes": [],
2590
+ "range_num_stages": [
2591
+ 0
2592
+ ],
2593
+ "range_multi_buffers": [
2594
+ null
2595
+ ],
2596
+ "range_flattens": [
2597
+ null
2598
+ ],
2599
+ "load_eviction_policies": [
2600
+ "",
2601
+ "",
2602
+ ""
2603
+ ],
2604
+ "num_warps": 4,
2605
+ "num_stages": 1,
2606
+ "indexing": [
2607
+ "pointer",
2608
+ "pointer",
2609
+ "pointer",
2610
+ "pointer"
2611
+ ],
2612
+ "pid_type": "flat"
2613
+ },
2614
+ "intermediate_4096_numtokens_48": {
2615
+ "block_sizes": [
2616
+ 32,
2617
+ 64
2618
+ ],
2619
+ "loop_orders": [
2620
+ [
2621
+ 0,
2622
+ 1
2623
+ ]
2624
+ ],
2625
+ "flatten_loops": [
2626
+ true
2627
+ ],
2628
+ "l2_groupings": [
2629
+ 1
2630
+ ],
2631
+ "range_unroll_factors": [
2632
+ 0
2633
+ ],
2634
+ "range_warp_specializes": [],
2635
+ "range_num_stages": [
2636
+ 0
2637
+ ],
2638
+ "range_multi_buffers": [
2639
+ null
2640
+ ],
2641
+ "range_flattens": [
2642
+ null
2643
+ ],
2644
+ "load_eviction_policies": [
2645
+ "",
2646
+ "",
2647
+ ""
2648
+ ],
2649
+ "num_warps": 8,
2650
+ "num_stages": 1,
2651
+ "indexing": [
2652
+ "pointer",
2653
+ "pointer",
2654
+ "pointer",
2655
+ "pointer"
2656
+ ],
2657
+ "pid_type": "flat"
2658
+ },
2659
+ "intermediate_8192_numtokens_48": {
2660
+ "block_sizes": [
2661
+ 1,
2662
+ 4096
2663
+ ],
2664
+ "loop_orders": [
2665
+ [
2666
+ 0,
2667
+ 1
2668
+ ]
2669
+ ],
2670
+ "flatten_loops": [
2671
+ true
2672
+ ],
2673
+ "l2_groupings": [
2674
+ 2
2675
+ ],
2676
+ "range_unroll_factors": [
2677
+ 0
2678
+ ],
2679
+ "range_warp_specializes": [],
2680
+ "range_num_stages": [
2681
+ 0
2682
+ ],
2683
+ "range_multi_buffers": [
2684
+ null
2685
+ ],
2686
+ "range_flattens": [
2687
+ null
2688
+ ],
2689
+ "load_eviction_policies": [
2690
+ "first",
2691
+ "",
2692
+ "last"
2693
+ ],
2694
+ "num_warps": 4,
2695
+ "num_stages": 2,
2696
+ "indexing": [
2697
+ "pointer",
2698
+ "tensor_descriptor",
2699
+ "tensor_descriptor",
2700
+ "pointer"
2701
+ ],
2702
+ "pid_type": "flat"
2703
+ },
2704
+ "intermediate_11008_numtokens_48": {
2705
+ "block_sizes": [
2706
+ 16,
2707
+ 256
2708
+ ],
2709
+ "loop_orders": [
2710
+ [
2711
+ 0,
2712
+ 1
2713
+ ]
2714
+ ],
2715
+ "flatten_loops": [
2716
+ true
2717
+ ],
2718
+ "l2_groupings": [
2719
+ 1
2720
+ ],
2721
+ "range_unroll_factors": [
2722
+ 0
2723
+ ],
2724
+ "range_warp_specializes": [],
2725
+ "range_num_stages": [
2726
+ 0
2727
+ ],
2728
+ "range_multi_buffers": [
2729
+ null
2730
+ ],
2731
+ "range_flattens": [
2732
+ null
2733
+ ],
2734
+ "load_eviction_policies": [
2735
+ "",
2736
+ "",
2737
+ ""
2738
+ ],
2739
+ "num_warps": 8,
2740
+ "num_stages": 1,
2741
+ "indexing": [
2742
+ "pointer",
2743
+ "pointer",
2744
+ "pointer",
2745
+ "pointer"
2746
+ ],
2747
+ "pid_type": "flat"
2748
+ },
2749
+ "intermediate_14336_numtokens_48": {
2750
+ "block_sizes": [
2751
+ 32,
2752
+ 256
2753
+ ],
2754
+ "loop_orders": [
2755
+ [
2756
+ 0,
2757
+ 1
2758
+ ]
2759
+ ],
2760
+ "flatten_loops": [
2761
+ true
2762
+ ],
2763
+ "l2_groupings": [
2764
+ 1
2765
+ ],
2766
+ "range_unroll_factors": [
2767
+ 0
2768
+ ],
2769
+ "range_warp_specializes": [],
2770
+ "range_num_stages": [
2771
+ 0
2772
+ ],
2773
+ "range_multi_buffers": [
2774
+ null
2775
+ ],
2776
+ "range_flattens": [
2777
+ null
2778
+ ],
2779
+ "load_eviction_policies": [
2780
+ "",
2781
+ "",
2782
+ ""
2783
+ ],
2784
+ "num_warps": 8,
2785
+ "num_stages": 1,
2786
+ "indexing": [
2787
+ "pointer",
2788
+ "pointer",
2789
+ "pointer",
2790
+ "pointer"
2791
+ ],
2792
+ "pid_type": "flat"
2793
+ },
2794
+ "intermediate_2048_numtokens_56": {
2795
+ "block_sizes": [
2796
+ 32,
2797
+ 32
2798
+ ],
2799
+ "loop_orders": [
2800
+ [
2801
+ 0,
2802
+ 1
2803
+ ]
2804
+ ],
2805
+ "flatten_loops": [
2806
+ true
2807
+ ],
2808
+ "l2_groupings": [
2809
+ 1
2810
+ ],
2811
+ "range_unroll_factors": [
2812
+ 0
2813
+ ],
2814
+ "range_warp_specializes": [],
2815
+ "range_num_stages": [
2816
+ 0
2817
+ ],
2818
+ "range_multi_buffers": [
2819
+ null
2820
+ ],
2821
+ "range_flattens": [
2822
+ null
2823
+ ],
2824
+ "load_eviction_policies": [
2825
+ "",
2826
+ "",
2827
+ ""
2828
+ ],
2829
+ "num_warps": 4,
2830
+ "num_stages": 1,
2831
+ "indexing": [
2832
+ "pointer",
2833
+ "pointer",
2834
+ "pointer",
2835
+ "pointer"
2836
+ ],
2837
+ "pid_type": "flat"
2838
+ },
2839
+ "intermediate_2880_numtokens_56": {
2840
+ "block_sizes": [
2841
+ 1,
2842
+ 2048
2843
+ ],
2844
+ "loop_orders": [
2845
+ [
2846
+ 0,
2847
+ 1
2848
+ ]
2849
+ ],
2850
+ "flatten_loops": [
2851
+ true
2852
+ ],
2853
+ "l2_groupings": [
2854
+ 1
2855
+ ],
2856
+ "range_unroll_factors": [
2857
+ 0
2858
+ ],
2859
+ "range_warp_specializes": [],
2860
+ "range_num_stages": [
2861
+ 0
2862
+ ],
2863
+ "range_multi_buffers": [
2864
+ null
2865
+ ],
2866
+ "range_flattens": [
2867
+ null
2868
+ ],
2869
+ "load_eviction_policies": [
2870
+ "",
2871
+ "",
2872
+ ""
2873
+ ],
2874
+ "num_warps": 8,
2875
+ "num_stages": 1,
2876
+ "indexing": [
2877
+ "pointer",
2878
+ "pointer",
2879
+ "pointer",
2880
+ "pointer"
2881
+ ],
2882
+ "pid_type": "flat"
2883
+ },
2884
+ "intermediate_4096_numtokens_56": {
2885
+ "block_sizes": [
2886
+ 32,
2887
+ 64
2888
+ ],
2889
+ "loop_orders": [
2890
+ [
2891
+ 0,
2892
+ 1
2893
+ ]
2894
+ ],
2895
+ "flatten_loops": [
2896
+ true
2897
+ ],
2898
+ "l2_groupings": [
2899
+ 1
2900
+ ],
2901
+ "range_unroll_factors": [
2902
+ 0
2903
+ ],
2904
+ "range_warp_specializes": [],
2905
+ "range_num_stages": [
2906
+ 0
2907
+ ],
2908
+ "range_multi_buffers": [
2909
+ null
2910
+ ],
2911
+ "range_flattens": [
2912
+ null
2913
+ ],
2914
+ "load_eviction_policies": [
2915
+ "",
2916
+ "",
2917
+ ""
2918
+ ],
2919
+ "num_warps": 8,
2920
+ "num_stages": 1,
2921
+ "indexing": [
2922
+ "pointer",
2923
+ "pointer",
2924
+ "pointer",
2925
+ "pointer"
2926
+ ],
2927
+ "pid_type": "flat"
2928
+ },
2929
+ "intermediate_8192_numtokens_56": {
2930
+ "block_sizes": [
2931
+ 32,
2932
+ 128
2933
+ ],
2934
+ "loop_orders": [
2935
+ [
2936
+ 0,
2937
+ 1
2938
+ ]
2939
+ ],
2940
+ "flatten_loops": [
2941
+ true
2942
+ ],
2943
+ "l2_groupings": [
2944
+ 1
2945
+ ],
2946
+ "range_unroll_factors": [
2947
+ 0
2948
+ ],
2949
+ "range_warp_specializes": [],
2950
+ "range_num_stages": [
2951
+ 0
2952
+ ],
2953
+ "range_multi_buffers": [
2954
+ null
2955
+ ],
2956
+ "range_flattens": [
2957
+ null
2958
+ ],
2959
+ "load_eviction_policies": [
2960
+ "",
2961
+ "",
2962
+ ""
2963
+ ],
2964
+ "num_warps": 8,
2965
+ "num_stages": 1,
2966
+ "indexing": [
2967
+ "pointer",
2968
+ "pointer",
2969
+ "pointer",
2970
+ "pointer"
2971
+ ],
2972
+ "pid_type": "flat"
2973
+ },
2974
+ "intermediate_11008_numtokens_56": {
2975
+ "block_sizes": [
2976
+ 1,
2977
+ 8192
2978
+ ],
2979
+ "loop_orders": [
2980
+ [
2981
+ 0,
2982
+ 1
2983
+ ]
2984
+ ],
2985
+ "flatten_loops": [
2986
+ true
2987
+ ],
2988
+ "l2_groupings": [
2989
+ 1
2990
+ ],
2991
+ "range_unroll_factors": [
2992
+ 0
2993
+ ],
2994
+ "range_warp_specializes": [],
2995
+ "range_num_stages": [
2996
+ 0
2997
+ ],
2998
+ "range_multi_buffers": [
2999
+ null
3000
+ ],
3001
+ "range_flattens": [
3002
+ null
3003
+ ],
3004
+ "load_eviction_policies": [
3005
+ "",
3006
+ "",
3007
+ ""
3008
+ ],
3009
+ "num_warps": 8,
3010
+ "num_stages": 1,
3011
+ "indexing": [
3012
+ "pointer",
3013
+ "pointer",
3014
+ "pointer",
3015
+ "pointer"
3016
+ ],
3017
+ "pid_type": "flat"
3018
+ },
3019
+ "intermediate_14336_numtokens_56": {
3020
+ "block_sizes": [
3021
+ 2,
3022
+ 4096
3023
+ ],
3024
+ "loop_orders": [
3025
+ [
3026
+ 0,
3027
+ 1
3028
+ ]
3029
+ ],
3030
+ "flatten_loops": [
3031
+ false
3032
+ ],
3033
+ "l2_groupings": [
3034
+ 32
3035
+ ],
3036
+ "range_unroll_factors": [
3037
+ 0
3038
+ ],
3039
+ "range_warp_specializes": [],
3040
+ "range_num_stages": [
3041
+ 0
3042
+ ],
3043
+ "range_multi_buffers": [
3044
+ null
3045
+ ],
3046
+ "range_flattens": [
3047
+ null
3048
+ ],
3049
+ "load_eviction_policies": [
3050
+ "first",
3051
+ "first",
3052
+ ""
3053
+ ],
3054
+ "num_warps": 2,
3055
+ "num_stages": 4,
3056
+ "indexing": [
3057
+ "tensor_descriptor",
3058
+ "pointer",
3059
+ "tensor_descriptor",
3060
+ "tensor_descriptor"
3061
+ ],
3062
+ "pid_type": "flat"
3063
+ },
3064
+ "intermediate_2048_numtokens_64": {
3065
+ "block_sizes": [
3066
+ 64,
3067
+ 32
3068
+ ],
3069
+ "loop_orders": [
3070
+ [
3071
+ 0,
3072
+ 1
3073
+ ]
3074
+ ],
3075
+ "flatten_loops": [
3076
+ true
3077
+ ],
3078
+ "l2_groupings": [
3079
+ 1
3080
+ ],
3081
+ "range_unroll_factors": [
3082
+ 0
3083
+ ],
3084
+ "range_warp_specializes": [],
3085
+ "range_num_stages": [
3086
+ 0
3087
+ ],
3088
+ "range_multi_buffers": [
3089
+ null
3090
+ ],
3091
+ "range_flattens": [
3092
+ null
3093
+ ],
3094
+ "load_eviction_policies": [
3095
+ "",
3096
+ "",
3097
+ ""
3098
+ ],
3099
+ "num_warps": 8,
3100
+ "num_stages": 1,
3101
+ "indexing": [
3102
+ "pointer",
3103
+ "pointer",
3104
+ "pointer",
3105
+ "pointer"
3106
+ ],
3107
+ "pid_type": "flat"
3108
+ },
3109
+ "intermediate_2880_numtokens_64": {
3110
+ "block_sizes": [
3111
+ 1,
3112
+ 2048
3113
+ ],
3114
+ "loop_orders": [
3115
+ [
3116
+ 0,
3117
+ 1
3118
+ ]
3119
+ ],
3120
+ "flatten_loops": [
3121
+ true
3122
+ ],
3123
+ "l2_groupings": [
3124
+ 1
3125
+ ],
3126
+ "range_unroll_factors": [
3127
+ 0
3128
+ ],
3129
+ "range_warp_specializes": [],
3130
+ "range_num_stages": [
3131
+ 0
3132
+ ],
3133
+ "range_multi_buffers": [
3134
+ null
3135
+ ],
3136
+ "range_flattens": [
3137
+ null
3138
+ ],
3139
+ "load_eviction_policies": [
3140
+ "",
3141
+ "",
3142
+ ""
3143
+ ],
3144
+ "num_warps": 8,
3145
+ "num_stages": 1,
3146
+ "indexing": [
3147
+ "pointer",
3148
+ "pointer",
3149
+ "pointer",
3150
+ "pointer"
3151
+ ],
3152
+ "pid_type": "flat"
3153
+ },
3154
+ "intermediate_4096_numtokens_64": {
3155
+ "block_sizes": [
3156
+ 64,
3157
+ 32
3158
+ ],
3159
+ "loop_orders": [
3160
+ [
3161
+ 0,
3162
+ 1
3163
+ ]
3164
+ ],
3165
+ "flatten_loops": [
3166
+ true
3167
+ ],
3168
+ "l2_groupings": [
3169
+ 1
3170
+ ],
3171
+ "range_unroll_factors": [
3172
+ 0
3173
+ ],
3174
+ "range_warp_specializes": [],
3175
+ "range_num_stages": [
3176
+ 0
3177
+ ],
3178
+ "range_multi_buffers": [
3179
+ null
3180
+ ],
3181
+ "range_flattens": [
3182
+ null
3183
+ ],
3184
+ "load_eviction_policies": [
3185
+ "",
3186
+ "",
3187
+ ""
3188
+ ],
3189
+ "num_warps": 8,
3190
+ "num_stages": 1,
3191
+ "indexing": [
3192
+ "pointer",
3193
+ "pointer",
3194
+ "pointer",
3195
+ "pointer"
3196
+ ],
3197
+ "pid_type": "flat"
3198
+ },
3199
+ "intermediate_8192_numtokens_64": {
3200
+ "block_sizes": [
3201
+ 64,
3202
+ 64
3203
+ ],
3204
+ "loop_orders": [
3205
+ [
3206
+ 0,
3207
+ 1
3208
+ ]
3209
+ ],
3210
+ "flatten_loops": [
3211
+ true
3212
+ ],
3213
+ "l2_groupings": [
3214
+ 1
3215
+ ],
3216
+ "range_unroll_factors": [
3217
+ 0
3218
+ ],
3219
+ "range_warp_specializes": [],
3220
+ "range_num_stages": [
3221
+ 0
3222
+ ],
3223
+ "range_multi_buffers": [
3224
+ null
3225
+ ],
3226
+ "range_flattens": [
3227
+ null
3228
+ ],
3229
+ "load_eviction_policies": [
3230
+ "",
3231
+ "",
3232
+ ""
3233
+ ],
3234
+ "num_warps": 8,
3235
+ "num_stages": 1,
3236
+ "indexing": [
3237
+ "pointer",
3238
+ "pointer",
3239
+ "pointer",
3240
+ "pointer"
3241
+ ],
3242
+ "pid_type": "flat"
3243
+ },
3244
+ "intermediate_11008_numtokens_64": {
3245
+ "block_sizes": [
3246
+ 1,
3247
+ 8192
3248
+ ],
3249
+ "loop_orders": [
3250
+ [
3251
+ 0,
3252
+ 1
3253
+ ]
3254
+ ],
3255
+ "flatten_loops": [
3256
+ true
3257
+ ],
3258
+ "l2_groupings": [
3259
+ 1
3260
+ ],
3261
+ "range_unroll_factors": [
3262
+ 0
3263
+ ],
3264
+ "range_warp_specializes": [],
3265
+ "range_num_stages": [
3266
+ 0
3267
+ ],
3268
+ "range_multi_buffers": [
3269
+ null
3270
+ ],
3271
+ "range_flattens": [
3272
+ null
3273
+ ],
3274
+ "load_eviction_policies": [
3275
+ "",
3276
+ "",
3277
+ ""
3278
+ ],
3279
+ "num_warps": 8,
3280
+ "num_stages": 1,
3281
+ "indexing": [
3282
+ "pointer",
3283
+ "pointer",
3284
+ "pointer",
3285
+ "pointer"
3286
+ ],
3287
+ "pid_type": "flat"
3288
+ },
3289
+ "intermediate_14336_numtokens_64": {
3290
+ "block_sizes": [
3291
+ 16,
3292
+ 512
3293
+ ],
3294
+ "loop_orders": [
3295
+ [
3296
+ 0,
3297
+ 1
3298
+ ]
3299
+ ],
3300
+ "flatten_loops": [
3301
+ true
3302
+ ],
3303
+ "l2_groupings": [
3304
+ 16
3305
+ ],
3306
+ "range_unroll_factors": [
3307
+ 0
3308
+ ],
3309
+ "range_warp_specializes": [],
3310
+ "range_num_stages": [
3311
+ 0
3312
+ ],
3313
+ "range_multi_buffers": [
3314
+ null
3315
+ ],
3316
+ "range_flattens": [
3317
+ null
3318
+ ],
3319
+ "load_eviction_policies": [
3320
+ "",
3321
+ "first",
3322
+ ""
3323
+ ],
3324
+ "num_warps": 4,
3325
+ "num_stages": 1,
3326
+ "indexing": [
3327
+ "pointer",
3328
+ "tensor_descriptor",
3329
+ "tensor_descriptor",
3330
+ "tensor_descriptor"
3331
+ ],
3332
+ "pid_type": "flat"
3333
+ },
3334
+ "intermediate_2048_numtokens_72": {
3335
+ "block_sizes": [
3336
+ 64,
3337
+ 32
3338
+ ],
3339
+ "loop_orders": [
3340
+ [
3341
+ 0,
3342
+ 1
3343
+ ]
3344
+ ],
3345
+ "flatten_loops": [
3346
+ true
3347
+ ],
3348
+ "l2_groupings": [
3349
+ 1
3350
+ ],
3351
+ "range_unroll_factors": [
3352
+ 0
3353
+ ],
3354
+ "range_warp_specializes": [],
3355
+ "range_num_stages": [
3356
+ 0
3357
+ ],
3358
+ "range_multi_buffers": [
3359
+ null
3360
+ ],
3361
+ "range_flattens": [
3362
+ null
3363
+ ],
3364
+ "load_eviction_policies": [
3365
+ "",
3366
+ "",
3367
+ ""
3368
+ ],
3369
+ "num_warps": 8,
3370
+ "num_stages": 1,
3371
+ "indexing": [
3372
+ "pointer",
3373
+ "pointer",
3374
+ "pointer",
3375
+ "pointer"
3376
+ ],
3377
+ "pid_type": "flat"
3378
+ },
3379
+ "intermediate_2880_numtokens_72": {
3380
+ "block_sizes": [
3381
+ 32,
3382
+ 64
3383
+ ],
3384
+ "loop_orders": [
3385
+ [
3386
+ 0,
3387
+ 1
3388
+ ]
3389
+ ],
3390
+ "flatten_loops": [
3391
+ true
3392
+ ],
3393
+ "l2_groupings": [
3394
+ 1
3395
+ ],
3396
+ "range_unroll_factors": [
3397
+ 0
3398
+ ],
3399
+ "range_warp_specializes": [],
3400
+ "range_num_stages": [
3401
+ 0
3402
+ ],
3403
+ "range_multi_buffers": [
3404
+ null
3405
+ ],
3406
+ "range_flattens": [
3407
+ null
3408
+ ],
3409
+ "load_eviction_policies": [
3410
+ "",
3411
+ "",
3412
+ ""
3413
+ ],
3414
+ "num_warps": 8,
3415
+ "num_stages": 1,
3416
+ "indexing": [
3417
+ "pointer",
3418
+ "pointer",
3419
+ "pointer",
3420
+ "pointer"
3421
+ ],
3422
+ "pid_type": "flat"
3423
+ },
3424
+ "intermediate_4096_numtokens_72": {
3425
+ "block_sizes": [
3426
+ 1,
3427
+ 1024
3428
+ ],
3429
+ "loop_orders": [
3430
+ [
3431
+ 0,
3432
+ 1
3433
+ ]
3434
+ ],
3435
+ "flatten_loops": [
3436
+ true
3437
+ ],
3438
+ "l2_groupings": [
3439
+ 4
3440
+ ],
3441
+ "range_unroll_factors": [
3442
+ 0
3443
+ ],
3444
+ "range_warp_specializes": [],
3445
+ "range_num_stages": [
3446
+ 0
3447
+ ],
3448
+ "range_multi_buffers": [
3449
+ null
3450
+ ],
3451
+ "range_flattens": [
3452
+ null
3453
+ ],
3454
+ "load_eviction_policies": [
3455
+ "",
3456
+ "",
3457
+ "first"
3458
+ ],
3459
+ "num_warps": 16,
3460
+ "num_stages": 2,
3461
+ "indexing": [
3462
+ "tensor_descriptor",
3463
+ "pointer",
3464
+ "pointer",
3465
+ "tensor_descriptor"
3466
+ ],
3467
+ "pid_type": "flat"
3468
+ },
3469
+ "intermediate_8192_numtokens_72": {
3470
+ "block_sizes": [
3471
+ 64,
3472
+ 128
3473
+ ],
3474
+ "loop_orders": [
3475
+ [
3476
+ 0,
3477
+ 1
3478
+ ]
3479
+ ],
3480
+ "flatten_loops": [
3481
+ true
3482
+ ],
3483
+ "l2_groupings": [
3484
+ 1
3485
+ ],
3486
+ "range_unroll_factors": [
3487
+ 0
3488
+ ],
3489
+ "range_warp_specializes": [],
3490
+ "range_num_stages": [
3491
+ 0
3492
+ ],
3493
+ "range_multi_buffers": [
3494
+ null
3495
+ ],
3496
+ "range_flattens": [
3497
+ null
3498
+ ],
3499
+ "load_eviction_policies": [
3500
+ "",
3501
+ "",
3502
+ ""
3503
+ ],
3504
+ "num_warps": 8,
3505
+ "num_stages": 1,
3506
+ "indexing": [
3507
+ "pointer",
3508
+ "pointer",
3509
+ "pointer",
3510
+ "pointer"
3511
+ ],
3512
+ "pid_type": "flat"
3513
+ },
3514
+ "intermediate_11008_numtokens_72": {
3515
+ "block_sizes": [
3516
+ 4,
3517
+ 256
3518
+ ],
3519
+ "loop_orders": [
3520
+ [
3521
+ 0,
3522
+ 1
3523
+ ]
3524
+ ],
3525
+ "flatten_loops": [
3526
+ true
3527
+ ],
3528
+ "l2_groupings": [
3529
+ 2
3530
+ ],
3531
+ "range_unroll_factors": [
3532
+ 0
3533
+ ],
3534
+ "range_warp_specializes": [],
3535
+ "range_num_stages": [
3536
+ 0
3537
+ ],
3538
+ "range_multi_buffers": [
3539
+ null
3540
+ ],
3541
+ "range_flattens": [
3542
+ null
3543
+ ],
3544
+ "load_eviction_policies": [
3545
+ "first",
3546
+ "last",
3547
+ "last"
3548
+ ],
3549
+ "num_warps": 32,
3550
+ "num_stages": 5,
3551
+ "indexing": [
3552
+ "pointer",
3553
+ "pointer",
3554
+ "tensor_descriptor",
3555
+ "tensor_descriptor"
3556
+ ],
3557
+ "pid_type": "flat"
3558
+ },
3559
+ "intermediate_14336_numtokens_72": {
3560
+ "block_sizes": [
3561
+ 32,
3562
+ 32
3563
+ ],
3564
+ "loop_orders": [
3565
+ [
3566
+ 0,
3567
+ 1
3568
+ ]
3569
+ ],
3570
+ "flatten_loops": [
3571
+ true
3572
+ ],
3573
+ "l2_groupings": [
3574
+ 2
3575
+ ],
3576
+ "range_unroll_factors": [
3577
+ 0
3578
+ ],
3579
+ "range_warp_specializes": [],
3580
+ "range_num_stages": [
3581
+ 0
3582
+ ],
3583
+ "range_multi_buffers": [
3584
+ null
3585
+ ],
3586
+ "range_flattens": [
3587
+ null
3588
+ ],
3589
+ "load_eviction_policies": [
3590
+ "last",
3591
+ "",
3592
+ "last"
3593
+ ],
3594
+ "num_warps": 8,
3595
+ "num_stages": 3,
3596
+ "indexing": [
3597
+ "pointer",
3598
+ "pointer",
3599
+ "tensor_descriptor",
3600
+ "pointer"
3601
+ ],
3602
+ "pid_type": "flat"
3603
+ },
3604
+ "intermediate_2048_numtokens_80": {
3605
+ "block_sizes": [
3606
+ 64,
3607
+ 32
3608
+ ],
3609
+ "loop_orders": [
3610
+ [
3611
+ 0,
3612
+ 1
3613
+ ]
3614
+ ],
3615
+ "flatten_loops": [
3616
+ true
3617
+ ],
3618
+ "l2_groupings": [
3619
+ 1
3620
+ ],
3621
+ "range_unroll_factors": [
3622
+ 0
3623
+ ],
3624
+ "range_warp_specializes": [],
3625
+ "range_num_stages": [
3626
+ 0
3627
+ ],
3628
+ "range_multi_buffers": [
3629
+ null
3630
+ ],
3631
+ "range_flattens": [
3632
+ null
3633
+ ],
3634
+ "load_eviction_policies": [
3635
+ "",
3636
+ "",
3637
+ ""
3638
+ ],
3639
+ "num_warps": 8,
3640
+ "num_stages": 1,
3641
+ "indexing": [
3642
+ "pointer",
3643
+ "pointer",
3644
+ "pointer",
3645
+ "pointer"
3646
+ ],
3647
+ "pid_type": "flat"
3648
+ },
3649
+ "intermediate_2880_numtokens_80": {
3650
+ "block_sizes": [
3651
+ 32,
3652
+ 64
3653
+ ],
3654
+ "loop_orders": [
3655
+ [
3656
+ 0,
3657
+ 1
3658
+ ]
3659
+ ],
3660
+ "flatten_loops": [
3661
+ true
3662
+ ],
3663
+ "l2_groupings": [
3664
+ 1
3665
+ ],
3666
+ "range_unroll_factors": [
3667
+ 0
3668
+ ],
3669
+ "range_warp_specializes": [],
3670
+ "range_num_stages": [
3671
+ 0
3672
+ ],
3673
+ "range_multi_buffers": [
3674
+ null
3675
+ ],
3676
+ "range_flattens": [
3677
+ null
3678
+ ],
3679
+ "load_eviction_policies": [
3680
+ "",
3681
+ "",
3682
+ ""
3683
+ ],
3684
+ "num_warps": 8,
3685
+ "num_stages": 1,
3686
+ "indexing": [
3687
+ "pointer",
3688
+ "pointer",
3689
+ "pointer",
3690
+ "pointer"
3691
+ ],
3692
+ "pid_type": "flat"
3693
+ },
3694
+ "intermediate_4096_numtokens_80": {
3695
+ "block_sizes": [
3696
+ 64,
3697
+ 64
3698
+ ],
3699
+ "loop_orders": [
3700
+ [
3701
+ 0,
3702
+ 1
3703
+ ]
3704
+ ],
3705
+ "flatten_loops": [
3706
+ true
3707
+ ],
3708
+ "l2_groupings": [
3709
+ 1
3710
+ ],
3711
+ "range_unroll_factors": [
3712
+ 0
3713
+ ],
3714
+ "range_warp_specializes": [],
3715
+ "range_num_stages": [
3716
+ 0
3717
+ ],
3718
+ "range_multi_buffers": [
3719
+ null
3720
+ ],
3721
+ "range_flattens": [
3722
+ null
3723
+ ],
3724
+ "load_eviction_policies": [
3725
+ "",
3726
+ "",
3727
+ ""
3728
+ ],
3729
+ "num_warps": 8,
3730
+ "num_stages": 1,
3731
+ "indexing": [
3732
+ "pointer",
3733
+ "pointer",
3734
+ "pointer",
3735
+ "pointer"
3736
+ ],
3737
+ "pid_type": "flat"
3738
+ },
3739
+ "intermediate_8192_numtokens_80": {
3740
+ "block_sizes": [
3741
+ 4,
3742
+ 128
3743
+ ],
3744
+ "loop_orders": [
3745
+ [
3746
+ 0,
3747
+ 1
3748
+ ]
3749
+ ],
3750
+ "flatten_loops": [
3751
+ true
3752
+ ],
3753
+ "l2_groupings": [
3754
+ 4
3755
+ ],
3756
+ "range_unroll_factors": [
3757
+ 0
3758
+ ],
3759
+ "range_warp_specializes": [],
3760
+ "range_num_stages": [
3761
+ 0
3762
+ ],
3763
+ "range_multi_buffers": [
3764
+ null
3765
+ ],
3766
+ "range_flattens": [
3767
+ null
3768
+ ],
3769
+ "load_eviction_policies": [
3770
+ "last",
3771
+ "last",
3772
+ "last"
3773
+ ],
3774
+ "num_warps": 16,
3775
+ "num_stages": 2,
3776
+ "indexing": [
3777
+ "pointer",
3778
+ "pointer",
3779
+ "pointer",
3780
+ "pointer"
3781
+ ],
3782
+ "pid_type": "flat"
3783
+ },
3784
+ "intermediate_11008_numtokens_80": {
3785
+ "block_sizes": [
3786
+ 1,
3787
+ 4096
3788
+ ],
3789
+ "loop_orders": [
3790
+ [
3791
+ 0,
3792
+ 1
3793
+ ]
3794
+ ],
3795
+ "flatten_loops": [
3796
+ true
3797
+ ],
3798
+ "l2_groupings": [
3799
+ 32
3800
+ ],
3801
+ "range_unroll_factors": [
3802
+ 0
3803
+ ],
3804
+ "range_warp_specializes": [],
3805
+ "range_num_stages": [
3806
+ 0
3807
+ ],
3808
+ "range_multi_buffers": [
3809
+ null
3810
+ ],
3811
+ "range_flattens": [
3812
+ null
3813
+ ],
3814
+ "load_eviction_policies": [
3815
+ "last",
3816
+ "last",
3817
+ ""
3818
+ ],
3819
+ "num_warps": 16,
3820
+ "num_stages": 2,
3821
+ "indexing": [
3822
+ "pointer",
3823
+ "tensor_descriptor",
3824
+ "tensor_descriptor",
3825
+ "tensor_descriptor"
3826
+ ],
3827
+ "pid_type": "flat"
3828
+ },
3829
+ "intermediate_14336_numtokens_80": {
3830
+ "block_sizes": [
3831
+ 2,
3832
+ 2048
3833
+ ],
3834
+ "loop_orders": [
3835
+ [
3836
+ 0,
3837
+ 1
3838
+ ]
3839
+ ],
3840
+ "flatten_loops": [
3841
+ true
3842
+ ],
3843
+ "l2_groupings": [
3844
+ 4
3845
+ ],
3846
+ "range_unroll_factors": [
3847
+ 0
3848
+ ],
3849
+ "range_warp_specializes": [],
3850
+ "range_num_stages": [
3851
+ 0
3852
+ ],
3853
+ "range_multi_buffers": [
3854
+ null
3855
+ ],
3856
+ "range_flattens": [
3857
+ null
3858
+ ],
3859
+ "load_eviction_policies": [
3860
+ "last",
3861
+ "last",
3862
+ ""
3863
+ ],
3864
+ "num_warps": 4,
3865
+ "num_stages": 6,
3866
+ "indexing": [
3867
+ "tensor_descriptor",
3868
+ "pointer",
3869
+ "pointer",
3870
+ "tensor_descriptor"
3871
+ ],
3872
+ "pid_type": "flat"
3873
+ },
3874
+ "intermediate_2048_numtokens_88": {
3875
+ "block_sizes": [
3876
+ 64,
3877
+ 32
3878
+ ],
3879
+ "loop_orders": [
3880
+ [
3881
+ 0,
3882
+ 1
3883
+ ]
3884
+ ],
3885
+ "flatten_loops": [
3886
+ true
3887
+ ],
3888
+ "l2_groupings": [
3889
+ 1
3890
+ ],
3891
+ "range_unroll_factors": [
3892
+ 0
3893
+ ],
3894
+ "range_warp_specializes": [],
3895
+ "range_num_stages": [
3896
+ 0
3897
+ ],
3898
+ "range_multi_buffers": [
3899
+ null
3900
+ ],
3901
+ "range_flattens": [
3902
+ null
3903
+ ],
3904
+ "load_eviction_policies": [
3905
+ "",
3906
+ "",
3907
+ ""
3908
+ ],
3909
+ "num_warps": 8,
3910
+ "num_stages": 1,
3911
+ "indexing": [
3912
+ "pointer",
3913
+ "pointer",
3914
+ "pointer",
3915
+ "pointer"
3916
+ ],
3917
+ "pid_type": "flat"
3918
+ },
3919
+ "intermediate_2880_numtokens_88": {
3920
+ "block_sizes": [
3921
+ 8,
3922
+ 256
3923
+ ],
3924
+ "loop_orders": [
3925
+ [
3926
+ 0,
3927
+ 1
3928
+ ]
3929
+ ],
3930
+ "flatten_loops": [
3931
+ true
3932
+ ],
3933
+ "l2_groupings": [
3934
+ 1
3935
+ ],
3936
+ "range_unroll_factors": [
3937
+ 0
3938
+ ],
3939
+ "range_warp_specializes": [],
3940
+ "range_num_stages": [
3941
+ 0
3942
+ ],
3943
+ "range_multi_buffers": [
3944
+ null
3945
+ ],
3946
+ "range_flattens": [
3947
+ null
3948
+ ],
3949
+ "load_eviction_policies": [
3950
+ "",
3951
+ "",
3952
+ ""
3953
+ ],
3954
+ "num_warps": 8,
3955
+ "num_stages": 1,
3956
+ "indexing": [
3957
+ "pointer",
3958
+ "pointer",
3959
+ "pointer",
3960
+ "pointer"
3961
+ ],
3962
+ "pid_type": "flat"
3963
+ },
3964
+ "intermediate_4096_numtokens_88": {
3965
+ "block_sizes": [
3966
+ 64,
3967
+ 64
3968
+ ],
3969
+ "loop_orders": [
3970
+ [
3971
+ 0,
3972
+ 1
3973
+ ]
3974
+ ],
3975
+ "flatten_loops": [
3976
+ true
3977
+ ],
3978
+ "l2_groupings": [
3979
+ 1
3980
+ ],
3981
+ "range_unroll_factors": [
3982
+ 0
3983
+ ],
3984
+ "range_warp_specializes": [],
3985
+ "range_num_stages": [
3986
+ 0
3987
+ ],
3988
+ "range_multi_buffers": [
3989
+ null
3990
+ ],
3991
+ "range_flattens": [
3992
+ null
3993
+ ],
3994
+ "load_eviction_policies": [
3995
+ "",
3996
+ "",
3997
+ ""
3998
+ ],
3999
+ "num_warps": 8,
4000
+ "num_stages": 1,
4001
+ "indexing": [
4002
+ "pointer",
4003
+ "pointer",
4004
+ "pointer",
4005
+ "pointer"
4006
+ ],
4007
+ "pid_type": "flat"
4008
+ },
4009
+ "intermediate_8192_numtokens_88": {
4010
+ "block_sizes": [
4011
+ 64,
4012
+ 128
4013
+ ],
4014
+ "loop_orders": [
4015
+ [
4016
+ 0,
4017
+ 1
4018
+ ]
4019
+ ],
4020
+ "flatten_loops": [
4021
+ true
4022
+ ],
4023
+ "l2_groupings": [
4024
+ 1
4025
+ ],
4026
+ "range_unroll_factors": [
4027
+ 0
4028
+ ],
4029
+ "range_warp_specializes": [],
4030
+ "range_num_stages": [
4031
+ 0
4032
+ ],
4033
+ "range_multi_buffers": [
4034
+ null
4035
+ ],
4036
+ "range_flattens": [
4037
+ null
4038
+ ],
4039
+ "load_eviction_policies": [
4040
+ "",
4041
+ "",
4042
+ ""
4043
+ ],
4044
+ "num_warps": 8,
4045
+ "num_stages": 1,
4046
+ "indexing": [
4047
+ "pointer",
4048
+ "pointer",
4049
+ "pointer",
4050
+ "pointer"
4051
+ ],
4052
+ "pid_type": "flat"
4053
+ },
4054
+ "intermediate_11008_numtokens_88": {
4055
+ "block_sizes": [
4056
+ 16,
4057
+ 2048
4058
+ ],
4059
+ "loop_orders": [
4060
+ [
4061
+ 0,
4062
+ 1
4063
+ ]
4064
+ ],
4065
+ "flatten_loops": [
4066
+ true
4067
+ ],
4068
+ "l2_groupings": [
4069
+ 2
4070
+ ],
4071
+ "range_unroll_factors": [
4072
+ 0
4073
+ ],
4074
+ "range_warp_specializes": [],
4075
+ "range_num_stages": [
4076
+ 0
4077
+ ],
4078
+ "range_multi_buffers": [
4079
+ null
4080
+ ],
4081
+ "range_flattens": [
4082
+ null
4083
+ ],
4084
+ "load_eviction_policies": [
4085
+ "first",
4086
+ "first",
4087
+ ""
4088
+ ],
4089
+ "num_warps": 32,
4090
+ "num_stages": 2,
4091
+ "indexing": [
4092
+ "tensor_descriptor",
4093
+ "tensor_descriptor",
4094
+ "pointer",
4095
+ "pointer"
4096
+ ],
4097
+ "pid_type": "flat"
4098
+ },
4099
+ "intermediate_14336_numtokens_88": {
4100
+ "block_sizes": [
4101
+ 4,
4102
+ 512
4103
+ ],
4104
+ "loop_orders": [
4105
+ [
4106
+ 0,
4107
+ 1
4108
+ ]
4109
+ ],
4110
+ "flatten_loops": [
4111
+ true
4112
+ ],
4113
+ "l2_groupings": [
4114
+ 16
4115
+ ],
4116
+ "range_unroll_factors": [
4117
+ 0
4118
+ ],
4119
+ "range_warp_specializes": [],
4120
+ "range_num_stages": [
4121
+ 0
4122
+ ],
4123
+ "range_multi_buffers": [
4124
+ null
4125
+ ],
4126
+ "range_flattens": [
4127
+ null
4128
+ ],
4129
+ "load_eviction_policies": [
4130
+ "",
4131
+ "",
4132
+ ""
4133
+ ],
4134
+ "num_warps": 32,
4135
+ "num_stages": 5,
4136
+ "indexing": [
4137
+ "pointer",
4138
+ "pointer",
4139
+ "pointer",
4140
+ "pointer"
4141
+ ],
4142
+ "pid_type": "flat"
4143
+ },
4144
+ "intermediate_2048_numtokens_96": {
4145
+ "block_sizes": [
4146
+ 64,
4147
+ 32
4148
+ ],
4149
+ "loop_orders": [
4150
+ [
4151
+ 0,
4152
+ 1
4153
+ ]
4154
+ ],
4155
+ "flatten_loops": [
4156
+ true
4157
+ ],
4158
+ "l2_groupings": [
4159
+ 1
4160
+ ],
4161
+ "range_unroll_factors": [
4162
+ 0
4163
+ ],
4164
+ "range_warp_specializes": [],
4165
+ "range_num_stages": [
4166
+ 0
4167
+ ],
4168
+ "range_multi_buffers": [
4169
+ null
4170
+ ],
4171
+ "range_flattens": [
4172
+ null
4173
+ ],
4174
+ "load_eviction_policies": [
4175
+ "",
4176
+ "",
4177
+ ""
4178
+ ],
4179
+ "num_warps": 8,
4180
+ "num_stages": 1,
4181
+ "indexing": [
4182
+ "pointer",
4183
+ "pointer",
4184
+ "pointer",
4185
+ "pointer"
4186
+ ],
4187
+ "pid_type": "flat"
4188
+ },
4189
+ "intermediate_2880_numtokens_96": {
4190
+ "block_sizes": [
4191
+ 32,
4192
+ 64
4193
+ ],
4194
+ "loop_orders": [
4195
+ [
4196
+ 0,
4197
+ 1
4198
+ ]
4199
+ ],
4200
+ "flatten_loops": [
4201
+ true
4202
+ ],
4203
+ "l2_groupings": [
4204
+ 1
4205
+ ],
4206
+ "range_unroll_factors": [
4207
+ 0
4208
+ ],
4209
+ "range_warp_specializes": [],
4210
+ "range_num_stages": [
4211
+ 0
4212
+ ],
4213
+ "range_multi_buffers": [
4214
+ null
4215
+ ],
4216
+ "range_flattens": [
4217
+ null
4218
+ ],
4219
+ "load_eviction_policies": [
4220
+ "",
4221
+ "",
4222
+ ""
4223
+ ],
4224
+ "num_warps": 8,
4225
+ "num_stages": 1,
4226
+ "indexing": [
4227
+ "pointer",
4228
+ "pointer",
4229
+ "pointer",
4230
+ "pointer"
4231
+ ],
4232
+ "pid_type": "flat"
4233
+ },
4234
+ "intermediate_4096_numtokens_96": {
4235
+ "block_sizes": [
4236
+ 1,
4237
+ 4096
4238
+ ],
4239
+ "loop_orders": [
4240
+ [
4241
+ 1,
4242
+ 0
4243
+ ]
4244
+ ],
4245
+ "flatten_loops": [
4246
+ false
4247
+ ],
4248
+ "l2_groupings": [
4249
+ 64
4250
+ ],
4251
+ "range_unroll_factors": [
4252
+ 0
4253
+ ],
4254
+ "range_warp_specializes": [],
4255
+ "range_num_stages": [
4256
+ 0
4257
+ ],
4258
+ "range_multi_buffers": [
4259
+ null
4260
+ ],
4261
+ "range_flattens": [
4262
+ null
4263
+ ],
4264
+ "load_eviction_policies": [
4265
+ "first",
4266
+ "last",
4267
+ "last"
4268
+ ],
4269
+ "num_warps": 4,
4270
+ "num_stages": 6,
4271
+ "indexing": [
4272
+ "pointer",
4273
+ "pointer",
4274
+ "tensor_descriptor",
4275
+ "pointer"
4276
+ ],
4277
+ "pid_type": "flat"
4278
+ },
4279
+ "intermediate_8192_numtokens_96": {
4280
+ "block_sizes": [
4281
+ 64,
4282
+ 128
4283
+ ],
4284
+ "loop_orders": [
4285
+ [
4286
+ 0,
4287
+ 1
4288
+ ]
4289
+ ],
4290
+ "flatten_loops": [
4291
+ true
4292
+ ],
4293
+ "l2_groupings": [
4294
+ 1
4295
+ ],
4296
+ "range_unroll_factors": [
4297
+ 0
4298
+ ],
4299
+ "range_warp_specializes": [],
4300
+ "range_num_stages": [
4301
+ 0
4302
+ ],
4303
+ "range_multi_buffers": [
4304
+ null
4305
+ ],
4306
+ "range_flattens": [
4307
+ null
4308
+ ],
4309
+ "load_eviction_policies": [
4310
+ "",
4311
+ "",
4312
+ ""
4313
+ ],
4314
+ "num_warps": 8,
4315
+ "num_stages": 1,
4316
+ "indexing": [
4317
+ "pointer",
4318
+ "pointer",
4319
+ "pointer",
4320
+ "pointer"
4321
+ ],
4322
+ "pid_type": "flat"
4323
+ },
4324
+ "intermediate_11008_numtokens_96": {
4325
+ "block_sizes": [
4326
+ 1,
4327
+ 2048
4328
+ ],
4329
+ "loop_orders": [
4330
+ [
4331
+ 1,
4332
+ 0
4333
+ ]
4334
+ ],
4335
+ "flatten_loops": [
4336
+ false
4337
+ ],
4338
+ "l2_groupings": [
4339
+ 64
4340
+ ],
4341
+ "range_unroll_factors": [
4342
+ 0
4343
+ ],
4344
+ "range_warp_specializes": [],
4345
+ "range_num_stages": [
4346
+ 0
4347
+ ],
4348
+ "range_multi_buffers": [
4349
+ null
4350
+ ],
4351
+ "range_flattens": [
4352
+ null
4353
+ ],
4354
+ "load_eviction_policies": [
4355
+ "first",
4356
+ "",
4357
+ ""
4358
+ ],
4359
+ "num_warps": 32,
4360
+ "num_stages": 2,
4361
+ "indexing": [
4362
+ "tensor_descriptor",
4363
+ "pointer",
4364
+ "pointer",
4365
+ "tensor_descriptor"
4366
+ ],
4367
+ "pid_type": "flat"
4368
+ },
4369
+ "intermediate_14336_numtokens_96": {
4370
+ "block_sizes": [
4371
+ 4,
4372
+ 4096
4373
+ ],
4374
+ "loop_orders": [
4375
+ [
4376
+ 0,
4377
+ 1
4378
+ ]
4379
+ ],
4380
+ "flatten_loops": [
4381
+ false
4382
+ ],
4383
+ "l2_groupings": [
4384
+ 32
4385
+ ],
4386
+ "range_unroll_factors": [
4387
+ 0
4388
+ ],
4389
+ "range_warp_specializes": [],
4390
+ "range_num_stages": [
4391
+ 0
4392
+ ],
4393
+ "range_multi_buffers": [
4394
+ null
4395
+ ],
4396
+ "range_flattens": [
4397
+ null
4398
+ ],
4399
+ "load_eviction_policies": [
4400
+ "last",
4401
+ "first",
4402
+ ""
4403
+ ],
4404
+ "num_warps": 4,
4405
+ "num_stages": 1,
4406
+ "indexing": [
4407
+ "tensor_descriptor",
4408
+ "tensor_descriptor",
4409
+ "pointer",
4410
+ "tensor_descriptor"
4411
+ ],
4412
+ "pid_type": "flat"
4413
+ },
4414
+ "intermediate_2048_numtokens_104": {
4415
+ "block_sizes": [
4416
+ 64,
4417
+ 32
4418
+ ],
4419
+ "loop_orders": [
4420
+ [
4421
+ 0,
4422
+ 1
4423
+ ]
4424
+ ],
4425
+ "flatten_loops": [
4426
+ true
4427
+ ],
4428
+ "l2_groupings": [
4429
+ 1
4430
+ ],
4431
+ "range_unroll_factors": [
4432
+ 0
4433
+ ],
4434
+ "range_warp_specializes": [],
4435
+ "range_num_stages": [
4436
+ 0
4437
+ ],
4438
+ "range_multi_buffers": [
4439
+ null
4440
+ ],
4441
+ "range_flattens": [
4442
+ null
4443
+ ],
4444
+ "load_eviction_policies": [
4445
+ "",
4446
+ "",
4447
+ ""
4448
+ ],
4449
+ "num_warps": 8,
4450
+ "num_stages": 1,
4451
+ "indexing": [
4452
+ "pointer",
4453
+ "pointer",
4454
+ "pointer",
4455
+ "pointer"
4456
+ ],
4457
+ "pid_type": "flat"
4458
+ },
4459
+ "intermediate_2880_numtokens_104": {
4460
+ "block_sizes": [
4461
+ 8,
4462
+ 256
4463
+ ],
4464
+ "loop_orders": [
4465
+ [
4466
+ 0,
4467
+ 1
4468
+ ]
4469
+ ],
4470
+ "flatten_loops": [
4471
+ true
4472
+ ],
4473
+ "l2_groupings": [
4474
+ 1
4475
+ ],
4476
+ "range_unroll_factors": [
4477
+ 0
4478
+ ],
4479
+ "range_warp_specializes": [],
4480
+ "range_num_stages": [
4481
+ 0
4482
+ ],
4483
+ "range_multi_buffers": [
4484
+ null
4485
+ ],
4486
+ "range_flattens": [
4487
+ null
4488
+ ],
4489
+ "load_eviction_policies": [
4490
+ "",
4491
+ "",
4492
+ ""
4493
+ ],
4494
+ "num_warps": 8,
4495
+ "num_stages": 1,
4496
+ "indexing": [
4497
+ "pointer",
4498
+ "pointer",
4499
+ "pointer",
4500
+ "pointer"
4501
+ ],
4502
+ "pid_type": "flat"
4503
+ },
4504
+ "intermediate_4096_numtokens_104": {
4505
+ "block_sizes": [
4506
+ 64,
4507
+ 64
4508
+ ],
4509
+ "loop_orders": [
4510
+ [
4511
+ 0,
4512
+ 1
4513
+ ]
4514
+ ],
4515
+ "flatten_loops": [
4516
+ true
4517
+ ],
4518
+ "l2_groupings": [
4519
+ 1
4520
+ ],
4521
+ "range_unroll_factors": [
4522
+ 0
4523
+ ],
4524
+ "range_warp_specializes": [],
4525
+ "range_num_stages": [
4526
+ 0
4527
+ ],
4528
+ "range_multi_buffers": [
4529
+ null
4530
+ ],
4531
+ "range_flattens": [
4532
+ null
4533
+ ],
4534
+ "load_eviction_policies": [
4535
+ "",
4536
+ "",
4537
+ ""
4538
+ ],
4539
+ "num_warps": 8,
4540
+ "num_stages": 1,
4541
+ "indexing": [
4542
+ "pointer",
4543
+ "pointer",
4544
+ "pointer",
4545
+ "pointer"
4546
+ ],
4547
+ "pid_type": "flat"
4548
+ },
4549
+ "intermediate_8192_numtokens_104": {
4550
+ "block_sizes": [
4551
+ 64,
4552
+ 128
4553
+ ],
4554
+ "loop_orders": [
4555
+ [
4556
+ 0,
4557
+ 1
4558
+ ]
4559
+ ],
4560
+ "flatten_loops": [
4561
+ true
4562
+ ],
4563
+ "l2_groupings": [
4564
+ 1
4565
+ ],
4566
+ "range_unroll_factors": [
4567
+ 0
4568
+ ],
4569
+ "range_warp_specializes": [],
4570
+ "range_num_stages": [
4571
+ 0
4572
+ ],
4573
+ "range_multi_buffers": [
4574
+ null
4575
+ ],
4576
+ "range_flattens": [
4577
+ null
4578
+ ],
4579
+ "load_eviction_policies": [
4580
+ "",
4581
+ "",
4582
+ ""
4583
+ ],
4584
+ "num_warps": 8,
4585
+ "num_stages": 1,
4586
+ "indexing": [
4587
+ "pointer",
4588
+ "pointer",
4589
+ "pointer",
4590
+ "pointer"
4591
+ ],
4592
+ "pid_type": "flat"
4593
+ },
4594
+ "intermediate_11008_numtokens_104": {
4595
+ "block_sizes": [
4596
+ 2,
4597
+ 8192
4598
+ ],
4599
+ "loop_orders": [
4600
+ [
4601
+ 0,
4602
+ 1
4603
+ ]
4604
+ ],
4605
+ "flatten_loops": [
4606
+ true
4607
+ ],
4608
+ "l2_groupings": [
4609
+ 1
4610
+ ],
4611
+ "range_unroll_factors": [
4612
+ 0
4613
+ ],
4614
+ "range_warp_specializes": [],
4615
+ "range_num_stages": [
4616
+ 0
4617
+ ],
4618
+ "range_multi_buffers": [
4619
+ null
4620
+ ],
4621
+ "range_flattens": [
4622
+ null
4623
+ ],
4624
+ "load_eviction_policies": [
4625
+ "",
4626
+ "last",
4627
+ "first"
4628
+ ],
4629
+ "num_warps": 8,
4630
+ "num_stages": 1,
4631
+ "indexing": [
4632
+ "pointer",
4633
+ "tensor_descriptor",
4634
+ "tensor_descriptor",
4635
+ "pointer"
4636
+ ],
4637
+ "pid_type": "flat"
4638
+ },
4639
+ "intermediate_14336_numtokens_104": {
4640
+ "block_sizes": [
4641
+ 8,
4642
+ 512
4643
+ ],
4644
+ "loop_orders": [
4645
+ [
4646
+ 0,
4647
+ 1
4648
+ ]
4649
+ ],
4650
+ "flatten_loops": [
4651
+ true
4652
+ ],
4653
+ "l2_groupings": [
4654
+ 2
4655
+ ],
4656
+ "range_unroll_factors": [
4657
+ 0
4658
+ ],
4659
+ "range_warp_specializes": [],
4660
+ "range_num_stages": [
4661
+ 0
4662
+ ],
4663
+ "range_multi_buffers": [
4664
+ null
4665
+ ],
4666
+ "range_flattens": [
4667
+ null
4668
+ ],
4669
+ "load_eviction_policies": [
4670
+ "last",
4671
+ "last",
4672
+ ""
4673
+ ],
4674
+ "num_warps": 8,
4675
+ "num_stages": 1,
4676
+ "indexing": [
4677
+ "pointer",
4678
+ "tensor_descriptor",
4679
+ "pointer",
4680
+ "tensor_descriptor"
4681
+ ],
4682
+ "pid_type": "flat"
4683
+ },
4684
+ "intermediate_2048_numtokens_112": {
4685
+ "block_sizes": [
4686
+ 64,
4687
+ 32
4688
+ ],
4689
+ "loop_orders": [
4690
+ [
4691
+ 0,
4692
+ 1
4693
+ ]
4694
+ ],
4695
+ "flatten_loops": [
4696
+ true
4697
+ ],
4698
+ "l2_groupings": [
4699
+ 1
4700
+ ],
4701
+ "range_unroll_factors": [
4702
+ 0
4703
+ ],
4704
+ "range_warp_specializes": [],
4705
+ "range_num_stages": [
4706
+ 0
4707
+ ],
4708
+ "range_multi_buffers": [
4709
+ null
4710
+ ],
4711
+ "range_flattens": [
4712
+ null
4713
+ ],
4714
+ "load_eviction_policies": [
4715
+ "",
4716
+ "",
4717
+ ""
4718
+ ],
4719
+ "num_warps": 8,
4720
+ "num_stages": 1,
4721
+ "indexing": [
4722
+ "pointer",
4723
+ "pointer",
4724
+ "pointer",
4725
+ "pointer"
4726
+ ],
4727
+ "pid_type": "flat"
4728
+ },
4729
+ "intermediate_2880_numtokens_112": {
4730
+ "block_sizes": [
4731
+ 2,
4732
+ 2048
4733
+ ],
4734
+ "loop_orders": [
4735
+ [
4736
+ 0,
4737
+ 1
4738
+ ]
4739
+ ],
4740
+ "flatten_loops": [
4741
+ true
4742
+ ],
4743
+ "l2_groupings": [
4744
+ 1
4745
+ ],
4746
+ "range_unroll_factors": [
4747
+ 0
4748
+ ],
4749
+ "range_warp_specializes": [],
4750
+ "range_num_stages": [
4751
+ 0
4752
+ ],
4753
+ "range_multi_buffers": [
4754
+ null
4755
+ ],
4756
+ "range_flattens": [
4757
+ null
4758
+ ],
4759
+ "load_eviction_policies": [
4760
+ "",
4761
+ "",
4762
+ ""
4763
+ ],
4764
+ "num_warps": 8,
4765
+ "num_stages": 1,
4766
+ "indexing": [
4767
+ "pointer",
4768
+ "pointer",
4769
+ "pointer",
4770
+ "pointer"
4771
+ ],
4772
+ "pid_type": "flat"
4773
+ },
4774
+ "intermediate_4096_numtokens_112": {
4775
+ "block_sizes": [
4776
+ 64,
4777
+ 64
4778
+ ],
4779
+ "loop_orders": [
4780
+ [
4781
+ 0,
4782
+ 1
4783
+ ]
4784
+ ],
4785
+ "flatten_loops": [
4786
+ true
4787
+ ],
4788
+ "l2_groupings": [
4789
+ 1
4790
+ ],
4791
+ "range_unroll_factors": [
4792
+ 0
4793
+ ],
4794
+ "range_warp_specializes": [],
4795
+ "range_num_stages": [
4796
+ 0
4797
+ ],
4798
+ "range_multi_buffers": [
4799
+ null
4800
+ ],
4801
+ "range_flattens": [
4802
+ null
4803
+ ],
4804
+ "load_eviction_policies": [
4805
+ "",
4806
+ "",
4807
+ ""
4808
+ ],
4809
+ "num_warps": 8,
4810
+ "num_stages": 1,
4811
+ "indexing": [
4812
+ "pointer",
4813
+ "pointer",
4814
+ "pointer",
4815
+ "pointer"
4816
+ ],
4817
+ "pid_type": "flat"
4818
+ },
4819
+ "intermediate_8192_numtokens_112": {
4820
+ "block_sizes": [
4821
+ 4,
4822
+ 512
4823
+ ],
4824
+ "loop_orders": [
4825
+ [
4826
+ 0,
4827
+ 1
4828
+ ]
4829
+ ],
4830
+ "flatten_loops": [
4831
+ true
4832
+ ],
4833
+ "l2_groupings": [
4834
+ 16
4835
+ ],
4836
+ "range_unroll_factors": [
4837
+ 0
4838
+ ],
4839
+ "range_warp_specializes": [],
4840
+ "range_num_stages": [
4841
+ 0
4842
+ ],
4843
+ "range_multi_buffers": [
4844
+ null
4845
+ ],
4846
+ "range_flattens": [
4847
+ null
4848
+ ],
4849
+ "load_eviction_policies": [
4850
+ "first",
4851
+ "last",
4852
+ "last"
4853
+ ],
4854
+ "num_warps": 16,
4855
+ "num_stages": 3,
4856
+ "indexing": [
4857
+ "pointer",
4858
+ "pointer",
4859
+ "tensor_descriptor",
4860
+ "pointer"
4861
+ ],
4862
+ "pid_type": "flat"
4863
+ },
4864
+ "intermediate_11008_numtokens_112": {
4865
+ "block_sizes": [
4866
+ 4,
4867
+ 1024
4868
+ ],
4869
+ "loop_orders": [
4870
+ [
4871
+ 0,
4872
+ 1
4873
+ ]
4874
+ ],
4875
+ "flatten_loops": [
4876
+ true
4877
+ ],
4878
+ "l2_groupings": [
4879
+ 2
4880
+ ],
4881
+ "range_unroll_factors": [
4882
+ 0
4883
+ ],
4884
+ "range_warp_specializes": [],
4885
+ "range_num_stages": [
4886
+ 0
4887
+ ],
4888
+ "range_multi_buffers": [
4889
+ null
4890
+ ],
4891
+ "range_flattens": [
4892
+ null
4893
+ ],
4894
+ "load_eviction_policies": [
4895
+ "",
4896
+ "last",
4897
+ "last"
4898
+ ],
4899
+ "num_warps": 32,
4900
+ "num_stages": 3,
4901
+ "indexing": [
4902
+ "pointer",
4903
+ "pointer",
4904
+ "tensor_descriptor",
4905
+ "tensor_descriptor"
4906
+ ],
4907
+ "pid_type": "flat"
4908
+ },
4909
+ "intermediate_14336_numtokens_112": {
4910
+ "block_sizes": [
4911
+ 64,
4912
+ 256
4913
+ ],
4914
+ "loop_orders": [
4915
+ [
4916
+ 0,
4917
+ 1
4918
+ ]
4919
+ ],
4920
+ "flatten_loops": [
4921
+ true
4922
+ ],
4923
+ "l2_groupings": [
4924
+ 1
4925
+ ],
4926
+ "range_unroll_factors": [
4927
+ 0
4928
+ ],
4929
+ "range_warp_specializes": [],
4930
+ "range_num_stages": [
4931
+ 0
4932
+ ],
4933
+ "range_multi_buffers": [
4934
+ null
4935
+ ],
4936
+ "range_flattens": [
4937
+ null
4938
+ ],
4939
+ "load_eviction_policies": [
4940
+ "",
4941
+ "",
4942
+ ""
4943
+ ],
4944
+ "num_warps": 8,
4945
+ "num_stages": 1,
4946
+ "indexing": [
4947
+ "pointer",
4948
+ "pointer",
4949
+ "pointer",
4950
+ "pointer"
4951
+ ],
4952
+ "pid_type": "flat"
4953
+ },
4954
+ "intermediate_2048_numtokens_120": {
4955
+ "block_sizes": [
4956
+ 8,
4957
+ 256
4958
+ ],
4959
+ "loop_orders": [
4960
+ [
4961
+ 0,
4962
+ 1
4963
+ ]
4964
+ ],
4965
+ "flatten_loops": [
4966
+ true
4967
+ ],
4968
+ "l2_groupings": [
4969
+ 16
4970
+ ],
4971
+ "range_unroll_factors": [
4972
+ 0
4973
+ ],
4974
+ "range_warp_specializes": [],
4975
+ "range_num_stages": [
4976
+ 0
4977
+ ],
4978
+ "range_multi_buffers": [
4979
+ null
4980
+ ],
4981
+ "range_flattens": [
4982
+ null
4983
+ ],
4984
+ "load_eviction_policies": [
4985
+ "first",
4986
+ "last",
4987
+ ""
4988
+ ],
4989
+ "num_warps": 4,
4990
+ "num_stages": 6,
4991
+ "indexing": [
4992
+ "tensor_descriptor",
4993
+ "tensor_descriptor",
4994
+ "pointer",
4995
+ "tensor_descriptor"
4996
+ ],
4997
+ "pid_type": "flat"
4998
+ },
4999
+ "intermediate_2880_numtokens_120": {
This file is too large to show in full.