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 16, 00:45 UTC
[FlashInfer] Upgrade to 0.2.0 (#11194)
Signed-off-by: Bowen Wang <abmfy@icloud.com> Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: youkaichao <youkaichao@gmail.com>
Bowen Wang committed
Jan 28, 2025 at 02:19 UTC
2bc3fbba0cf5b07fabb798d41b153b895d30c7b4
10 files changed
+257
-75
.buildkite/test-pipeline.yaml
+10
-1
@@ -183,7 +183,16 @@ steps:
183
- vllm/
184
- tests/v1
185
commands:
186
- - VLLM_USE_V1=1 pytest -v -s v1
186
+ # split the test to avoid interference
187
+ - VLLM_USE_V1=1 pytest -v -s v1/core
188
+ - VLLM_USE_V1=1 pytest -v -s v1/engine
189
+ - VLLM_USE_V1=1 pytest -v -s v1/sample
190
+ - VLLM_USE_V1=1 pytest -v -s v1/worker
191
+ - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py
192
+ - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py
193
+ # TODO: accuracy does not match, whether setting
194
+ # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
195
+ - VLLM_USE_V1=1 pytest -v -s v1/e2e
196
197
- label: Examples Test # 25min
198
working_dir: "/vllm-workspace/examples"
Dockerfile
+21
-2
@@ -149,7 +149,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
149
150
#################### vLLM installation IMAGE ####################
151
# image with vLLM installed
152
-FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
152
+# TODO: Restore to base image after FlashInfer AOT wheel fixed
153
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base
154
ARG CUDA_VERSION=12.4.1
155
ARG PYTHON_VERSION=3.12
156
WORKDIR /vllm-workspace
@@ -194,12 +195,30 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
195
--mount=type=cache,target=/root/.cache/pip \
196
python3 -m pip install dist/*.whl --verbose
197
198
+# How to build this FlashInfer wheel:
199
+# $ export FLASHINFER_ENABLE_AOT=1
200
+# $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+
201
+# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX'
202
+# $ git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
203
+# $ cd flashinfer
204
+# $ git checkout 524304395bd1d8cd7d07db083859523fcaa246a4
205
+# $ python3 setup.py bdist_wheel --dist-dir=dist --verbose
206
+
207
RUN --mount=type=cache,target=/root/.cache/pip \
208
. /etc/environment && \
209
if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
200
- python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
210
+ python3 -m pip install https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.0.post1-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
211
fi
212
COPY examples examples
213
+
214
+# Although we build Flashinfer with AOT mode, there's still
215
+# some issues w.r.t. JIT compilation. Therefore we need to
216
+# install build dependencies for JIT compilation.
217
+# TODO: Remove this once FlashInfer AOT wheel is fixed
218
+COPY requirements-build.txt requirements-build.txt
219
+RUN --mount=type=cache,target=/root/.cache/pip \
220
+ python3 -m pip install -r requirements-build.txt
221
+
222
#################### vLLM installation IMAGE ####################
223
224
#################### TEST IMAGE ####################
tests/basic_correctness/test_basic_correctness.py
+3
-2
@@ -61,9 +61,10 @@ def test_models(
61
if backend == "FLASHINFER" and current_platform.is_rocm():
62
pytest.skip("Flashinfer does not support ROCm/HIP.")
63
64
- if backend == "XFORMERS" and model == "google/gemma-2-2b-it":
64
+ if backend in ("XFORMERS",
65
+ "FLASHINFER") and model == "google/gemma-2-2b-it":
66
pytest.skip(
66
- "XFORMERS does not support gemma2 with full context length.")
67
+ f"{backend} does not support gemma2 with full context length.")
68
69
os.environ["VLLM_ATTENTION_BACKEND"] = backend
70
tests/compile/test_basic_correctness.py
+1
-1
@@ -58,7 +58,7 @@ test_settings = [
58
model_args=["--task", "embed"],
59
pp_size=1,
60
tp_size=1,
61
- attn_backend="FLASHINFER",
61
+ attn_backend="FLASH_ATTN",
62
method="encode",
63
fullgraph=True,
64
),
tests/kernels/test_flashinfer.py
+37
-37
@@ -133,17 +133,19 @@ def test_flashinfer_decode_with_paged_kv(
133
use_tensor_cores=(
134
(num_query_heads//num_kv_heads) > 4)
135
)
136
- wrapper.begin_forward(kv_indptr,
137
- kv_indices,
138
- kv_last_page_lens,
139
- num_query_heads,
140
- num_kv_heads,
141
- head_size,
142
- block_size,
143
- "NONE",
144
- data_type=dtype)
145
-
146
- output = wrapper.forward(query, key_value_cache, logits_soft_cap=soft_cap)
136
+ wrapper.plan(kv_indptr,
137
+ kv_indices,
138
+ kv_last_page_lens,
139
+ num_query_heads,
140
+ num_kv_heads,
141
+ head_size,
142
+ block_size,
143
+ "NONE",
144
+ q_data_type=dtype,
145
+ kv_data_type=dtype,
146
+ logits_soft_cap=soft_cap)
147
+
148
+ output = wrapper.run(query, key_value_cache)
149
150
ref_output = ref_paged_attn(query=query,
151
key_cache=key_cache,
@@ -228,7 +230,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
230
workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
231
wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
232
workspace_buffer, "NHD")
231
- wrapper.begin_forward(
233
+ wrapper.plan(
234
qo_indptr,
235
kv_indptr,
236
kv_indices,
@@ -237,12 +239,14 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
239
num_kv_heads,
240
head_size,
241
block_size,
242
+ q_data_type=dtype,
243
+ kv_data_type=dtype,
244
+ logits_soft_cap=soft_cap,
245
)
246
242
- output = wrapper.forward(
247
+ output = wrapper.run(
248
query,
249
key_value_cache,
245
- logits_soft_cap=soft_cap,
250
)
251
252
ref_output = ref_paged_attn(query=query,
@@ -253,7 +257,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
257
block_tables=block_tables,
258
scale=scale,
259
soft_cap=soft_cap)
256
- torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
260
+ torch.testing.assert_close(output, ref_output, atol=5e-2, rtol=1e-2), \
261
f"{torch.max(torch.abs(output - ref_output))}"
262
263
@@ -332,7 +336,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
336
workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.int8)
337
wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
338
workspace_buffer, "NHD")
335
- wrapper.begin_forward(
339
+ wrapper.plan(
340
qo_indptr,
341
kv_indptr,
342
kv_indices,
@@ -341,13 +345,12 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
345
num_kv_heads,
346
head_size,
347
block_size,
348
+ q_data_type=dtype,
349
+ kv_data_type=kv_cache_dtype,
350
+ logits_soft_cap=soft_cap,
351
)
352
346
- output = wrapper.forward(query,
347
- kv_cache_fp8,
348
- logits_soft_cap=soft_cap,
349
- k_scale=k_scale,
350
- v_scale=v_scale)
353
+ output = wrapper.run(query, kv_cache_fp8, k_scale=k_scale, v_scale=v_scale)
354
355
ref_output = ref_paged_attn(query=query,
356
key_cache=key_cache.squeeze(1),
@@ -360,7 +363,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
363
del query
364
del block_tables
365
# verify prefill fp8
363
- torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
366
+ torch.testing.assert_close(output, ref_output, atol=5e-2, rtol=1e-2), \
367
f"{torch.max(torch.abs(output - ref_output))}"
368
369
@@ -439,21 +442,18 @@ def test_flashinfer_decode_with_paged_fp8_kv(
442
wrapper = flashinfer.\
443
BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD",
444
use_tensor_cores=use_tensor_cores)
442
- wrapper.begin_forward(kv_indptr,
443
- kv_indices,
444
- kv_last_page_lens,
445
- num_query_heads,
446
- num_kv_heads,
447
- head_size,
448
- block_size,
449
- "NONE",
450
- data_type=dtype,
451
- q_data_type=dtype)
452
- output = wrapper.forward(query,
453
- kv_cache_fp8,
454
- logits_soft_cap=soft_cap,
455
- k_scale=k_scale,
456
- v_scale=v_scale)
445
+ wrapper.plan(kv_indptr,
446
+ kv_indices,
447
+ kv_last_page_lens,
448
+ num_query_heads,
449
+ num_kv_heads,
450
+ head_size,
451
+ block_size,
452
+ "NONE",
453
+ q_data_type=dtype,
454
+ kv_data_type=kv_cache_dtype,
455
+ logits_soft_cap=soft_cap)
456
+ output = wrapper.run(query, kv_cache_fp8, k_scale=k_scale, v_scale=v_scale)
457
key_cache = key_value_cache[:, 0, :, :, :].squeeze(1)
458
value_cache = key_value_cache[:, 1, :, :, :].squeeze(1)
459
vllm/attention/backends/flashinfer.py
+162
-21
@@ -1,3 +1,4 @@
1
+import dataclasses
2
from collections import defaultdict
3
from contextlib import contextmanager
4
from dataclasses import dataclass
@@ -13,9 +14,11 @@ try:
14
from vllm.vllm_flash_attn import flash_attn_varlen_func
15
FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024
16
except ImportError:
16
- BatchDecodeWithPagedKVCacheWrapper = None
17
- CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None
18
- BatchPrefillWithPagedKVCacheWrapper = None
17
+ # Avoid turning these types into variables during type checking
18
+ if not TYPE_CHECKING:
19
+ BatchDecodeWithPagedKVCacheWrapper = None
20
+ CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None
21
+ BatchPrefillWithPagedKVCacheWrapper = None
22
FLASHINFER_WORKSPACE_BUFFER_SIZE = 0
23
24
import torch
@@ -30,7 +33,9 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
33
from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping,
34
compute_slot_mapping_start_idx,
35
is_block_tables_empty)
36
+from vllm.attention.layer import Attention
37
from vllm.attention.ops.paged_attn import PagedAttention
38
+from vllm.config import VllmConfig, get_current_vllm_config
39
from vllm.utils import (async_tensor_h2d, get_kv_cache_torch_dtype,
40
make_tensor_with_pad)
41
@@ -99,6 +104,72 @@ class FlashInferBackend(AttentionBackend):
104
raise ValueError(f"Unrecognized FP8 dtype: {kv_cache_dtype}")
105
106
107
+@dataclass
108
+class PerLayerParameters:
109
+ """
110
+ Currently, FlashInfer backend only support models in which all layers share
111
+ the same values for the following hyperparameters.
112
+ """
113
+
114
+ window_left: int
115
+ logits_soft_cap: Optional[float]
116
+ sm_scale: float
117
+
118
+
119
+def get_per_layer_parameters(
120
+ vllm_config: VllmConfig) -> Dict[str, PerLayerParameters]:
121
+ """
122
+ Scan all attention layers and determine some hyperparameters
123
+ to use during `plan`.
124
+ """
125
+
126
+ layers = vllm_config.compilation_config.static_forward_context
127
+ per_layer_params: Dict[str, PerLayerParameters] = {}
128
+
129
+ for key, layer in layers.items():
130
+ assert isinstance(layer, Attention)
131
+
132
+ impl = layer.impl
133
+ assert isinstance(impl, FlashInferImpl)
134
+
135
+ # Infer hyperparameters from the attention layer
136
+ window_size = impl.sliding_window
137
+ window_left = window_size[0] if window_size is not None else -1
138
+ logits_soft_cap = impl.logits_soft_cap
139
+ sm_scale = impl.scale
140
+
141
+ per_layer_params[key] = PerLayerParameters(window_left,
142
+ logits_soft_cap, sm_scale)
143
+
144
+ return per_layer_params
145
+
146
+
147
+def infer_global_hyperparameters(
148
+ per_layer_params: Dict[str, PerLayerParameters]) -> PerLayerParameters:
149
+ """
150
+ Currently, FlashInfer backend only support models in which all layers share
151
+ the same values for the following hyperparameters:
152
+ - `window_left`
153
+ - `logits_soft_cap`
154
+ - `sm_scale`
155
+
156
+ So this function asserts that all layers share the same values for these
157
+ hyperparameters and returns the global values.
158
+ """
159
+
160
+ assert len(per_layer_params) > 0, "No attention layers found in the model."
161
+
162
+ param_sets = list(per_layer_params.values())
163
+ global_params = param_sets[0]
164
+ for params in param_sets:
165
+ assert params == global_params, (
166
+ "FlashInfer backend currently only supports models in which all "
167
+ "layers share the same values for the following hyperparameters: "
168
+ "`window_left`, `logits_soft_cap`, `sm_scale`.")
169
+
170
+ return global_params
171
+
172
+
173
class FlashInferState(AttentionState):
174
175
def __init__(self, runner):
@@ -108,6 +179,11 @@ class FlashInferState(AttentionState):
179
self._decode_wrapper = None
180
self._prefill_wrapper = None
181
182
+ # Global hyperparameters shared by all attention layers
183
+ self.global_hyperparameters: Optional[PerLayerParameters] = None
184
+
185
+ self.vllm_config = get_current_vllm_config()
186
+
187
def _get_workspace_buffer(self):
188
if self._workspace_buffer is None:
189
self._workspace_buffer = torch.empty(
@@ -215,6 +291,9 @@ class FlashInferState(AttentionState):
291
batch_size + 1,
292
dtype=torch.int32)
293
294
+ global_params = infer_global_hyperparameters(
295
+ get_per_layer_parameters(self.vllm_config))
296
+
297
attn_metadata = self.runner.attn_backend.make_metadata(
298
num_prefills=0,
299
slot_mapping=self._graph_slot_mapping[:batch_size],
@@ -238,7 +317,9 @@ class FlashInferState(AttentionState):
317
q_data_type=self.runner.model_config.dtype,
318
use_cuda_graph=True,
319
decode_wrapper=self._graph_decode_wrapper,
241
- prefill_wrapper=None)
320
+ prefill_wrapper=None,
321
+ **dataclasses.asdict(global_params),
322
+ )
323
attn_metadata.begin_forward()
324
return attn_metadata
325
@@ -325,9 +406,28 @@ class FlashInferMetadata(AttentionMetadata):
406
data_type: torch.dtype = None
407
# The data type of the query
408
q_data_type: torch.dtype = None
328
- device: torch.device = torch.device("cuda")
409
+ # FlashInfer 0.2 encourages passing host tensors
410
+ device: torch.device = torch.device("cpu")
411
is_profile_run: bool = False
412
413
+ # The FlashInfer backend currently supports only models in which all layers
414
+ # share the same following hyperparameters:
415
+
416
+ # The left (inclusive) window size for the attention window, when
417
+ # set to `-1`, the window size will be set to the full length of
418
+ # the sequence. Defaults to `-1`.
419
+ window_left: int = -1
420
+ # The attention logits soft capping value (used in Gemini, Grok and
421
+ # Gemma-2, etc.), if not provided, will be set to `0`. If greater
422
+ # than 0, the logits will be capped according to formula:
423
+ # $$\texttt{logits\_soft\_cap} \times
424
+ # \mathrm{tanh}(x / \texttt{logits\_soft\_cap})$$,
425
+ # where $x$ is the input logits.
426
+ logits_soft_cap: Optional[float] = None
427
+ # The scale used in softmax, if not provided, will be set to
428
+ # `1.0 / sqrt(head_dim)`.
429
+ sm_scale: Optional[float] = None
430
+
431
def __post_init__(self):
432
# Refer to
433
# https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
@@ -363,14 +463,21 @@ class FlashInferMetadata(AttentionMetadata):
463
self.block_table_bound = self.block_table_bound.to(self.device)
464
self.seq_lens_tensor = self.seq_lens_tensor.to(self.device)
465
self.paged_kv_indices = self.paged_kv_indices.to(self.device)
366
- self.prefill_wrapper.end_forward()
367
- self.prefill_wrapper.begin_forward(
466
+ self.prefill_wrapper.plan(
467
self.query_start_loc,
468
self.paged_kv_indptr[:self.num_prefills + 1],
469
self.paged_kv_indices,
470
self.paged_kv_last_page_len[:self.num_prefills],
372
- self.num_qo_heads, self.num_kv_heads, self.head_dim,
373
- self.page_size)
471
+ self.num_qo_heads,
472
+ self.num_kv_heads,
473
+ self.head_dim,
474
+ self.page_size,
475
+ causal=True,
476
+ sm_scale=self.sm_scale,
477
+ window_left=self.window_left,
478
+ logits_soft_cap=self.logits_soft_cap,
479
+ q_data_type=self.q_data_type,
480
+ kv_data_type=self.data_type)
481
if self.num_decode_tokens > 0:
482
assert self.paged_kv_indices is not None
483
assert self.paged_kv_indptr is not None
@@ -386,8 +493,7 @@ class FlashInferMetadata(AttentionMetadata):
493
self.seq_lens_tensor = self.seq_lens_tensor.to(self.device)
494
495
assert self.decode_wrapper is not None
389
- self.decode_wrapper.end_forward()
390
- self.decode_wrapper.begin_forward(
496
+ self.decode_wrapper.plan(
497
self.paged_kv_indptr[self.num_prefills:],
498
self.paged_kv_indices,
499
self.paged_kv_last_page_len[self.num_prefills:],
@@ -397,8 +503,11 @@ class FlashInferMetadata(AttentionMetadata):
503
self.page_size,
504
# Disable flashinfer's pos encoding and use vllm's rope.
505
pos_encoding_mode="NONE",
506
+ window_left=self.window_left,
507
+ logits_soft_cap=self.logits_soft_cap,
508
+ sm_scale=self.sm_scale,
509
# kv-cache data type.
401
- data_type=self.data_type,
510
+ kv_data_type=self.data_type,
511
# query data type.
512
q_data_type=self.q_data_type)
513
@@ -496,6 +605,11 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
605
self.sliding_window = input_builder.sliding_window
606
self.block_size = input_builder.block_size
607
608
+ # Global hyperparameters shared by all attention layers
609
+ self.global_hyperparameters: Optional[PerLayerParameters] = None
610
+
611
+ self.vllm_config = get_current_vllm_config()
612
+
613
def prepare(self):
614
self.slot_mapping: List[int] = []
615
self.prefill_seq_lens: List[int] = []
@@ -528,6 +642,20 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
642
self.total_blocks = 0
643
self.is_profile_run: bool = False
644
645
+ if self.global_hyperparameters is None:
646
+ # Infer global hyperparameters, since currently we only support
647
+ # models in which all layers share the same values for the
648
+ # following hyperparameters:
649
+ # - `window_left`
650
+ # - `logits_soft_cap`
651
+ # - `sm_scale`
652
+ inferred_params = infer_global_hyperparameters(
653
+ get_per_layer_parameters(self.vllm_config))
654
+ self.global_hyperparameters = inferred_params
655
+ self.window_left = inferred_params.window_left
656
+ self.logits_soft_cap = inferred_params.logits_soft_cap
657
+ self.sm_scale = inferred_params.sm_scale
658
+
659
def _add_seq_group(
660
self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
661
chunked_prefill_enabled: bool):
@@ -756,7 +884,11 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
884
data_type=kv_cache_dtype,
885
q_data_type=self.runner.model_config.dtype,
886
use_cuda_graph=use_captured_graph,
759
- is_profile_run=self.is_profile_run)
887
+ is_profile_run=self.is_profile_run,
888
+ window_left=self.window_left,
889
+ logits_soft_cap=self.logits_soft_cap,
890
+ sm_scale=self.sm_scale,
891
+ )
892
893
894
class FlashInferImpl(AttentionImpl):
@@ -885,25 +1017,34 @@ class FlashInferImpl(AttentionImpl):
1017
else:
1018
assert prefill_meta is not None
1019
assert prefill_meta.prefill_wrapper is not None
888
- prefill_output = prefill_meta.prefill_wrapper.forward(
1020
+
1021
+ assert prefill_meta.prefill_wrapper._causal
1022
+ assert prefill_meta.prefill_wrapper._window_left == window_left
1023
+ assert prefill_meta.prefill_wrapper._logits_soft_cap == (
1024
+ logits_soft_cap or 0.0)
1025
+ assert prefill_meta.prefill_wrapper._sm_scale == softmax_scale
1026
+
1027
+ prefill_output = prefill_meta.prefill_wrapper.run(
1028
query,
1029
kv_cache,
891
- logits_soft_cap=logits_soft_cap,
892
- causal=True,
1030
k_scale=layer._k_scale_float,
1031
v_scale=layer._v_scale_float,
895
- window_left=window_left)
1032
+ )
1033
if decode_meta := attn_metadata.decode_metadata:
1034
assert decode_meta is not None
1035
assert decode_meta.decode_wrapper is not None
899
- decode_output = decode_meta.decode_wrapper.forward(
1036
+
1037
+ assert decode_meta.decode_wrapper._window_left == window_left
1038
+ assert decode_meta.decode_wrapper._logits_soft_cap == (
1039
+ logits_soft_cap or 0.0)
1040
+ assert decode_meta.decode_wrapper._sm_scale == softmax_scale
1041
+
1042
+ decode_output = decode_meta.decode_wrapper.run(
1043
decode_query,
1044
kv_cache,
902
- sm_scale=softmax_scale,
903
- logits_soft_cap=logits_soft_cap,
1045
k_scale=layer._k_scale_float,
1046
v_scale=layer._v_scale_float,
906
- window_left=window_left)
1047
+ )
1048
1049
if prefill_output is None and decode_output is not None:
1050
# Decode only batch.
vllm/config.py
+6
-4
@@ -310,14 +310,15 @@ class ModelConfig:
310
(self.hf_text_config.model_type in ["gemma2", "cohere2"]))
311
312
if (not self.disable_sliding_window and has_interleaved_attention):
313
- if envs.VLLM_ATTENTION_BACKEND == "XFORMERS":
313
+ if (backend :=
314
+ envs.VLLM_ATTENTION_BACKEND) in ("XFORMERS", "FLASHINFER"):
315
sliding_window_len_min = get_min_sliding_window(
316
self.hf_text_config.sliding_window)
317
318
logger.warning_once(
319
f"{self.hf_text_config.model_type} has interleaved "
320
"attention, which is currently not supported by the "
320
- "XFORMERS backend. Disabling sliding window and capping "
321
+ f"{backend} backend. Disabling sliding window and capping "
322
"the max length to the sliding window size "
323
f"({sliding_window_len_min}).")
324
self.disable_sliding_window = True
@@ -3310,7 +3311,7 @@ _current_vllm_config: Optional[VllmConfig] = None
3311
3312
3313
@contextmanager
3313
-def set_current_vllm_config(vllm_config: VllmConfig):
3314
+def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):
3315
"""
3316
Temporarily set the current VLLM config.
3317
Used during model initialization.
@@ -3330,7 +3331,8 @@ def set_current_vllm_config(vllm_config: VllmConfig):
3331
vllm_config.compilation_config.enabled_custom_ops)
3332
logger.debug("disabled custom ops: %s",
3333
vllm_config.compilation_config.disabled_custom_ops)
3333
- if vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3334
+ if check_compile and \
3335
+ vllm_config.compilation_config.level == CompilationLevel.PIECEWISE \
3336
and compilation_counter.num_models_seen == num_models_seen:
3337
# If the model supports compilation,
3338
# compilation_counter.num_models_seen should be increased
vllm/model_executor/model_loader/loader.py
+2
-2
@@ -114,7 +114,7 @@ def _initialize_model(
114
all_params = [param.name for param in signatures.parameters.values()]
115
if "vllm_config" in all_params and "prefix" in all_params:
116
# new-style model class
117
- with set_current_vllm_config(vllm_config):
117
+ with set_current_vllm_config(vllm_config, check_compile=True):
118
return model_class(vllm_config=vllm_config, prefix=prefix)
119
120
msg = ("vLLM model class should accept `vllm_config` and `prefix` as "
@@ -142,7 +142,7 @@ def _initialize_model(
142
kwargs["lora_config"] = vllm_config.lora_config
143
if "scheduler_config" in all_params:
144
kwargs["scheduler_config"] = vllm_config.scheduler_config
145
- with set_current_vllm_config(vllm_config):
145
+ with set_current_vllm_config(vllm_config, check_compile=True):
146
return model_class(**kwargs)
147
148
vllm/model_executor/model_loader/tensorizer.py
+2
-1
@@ -288,7 +288,8 @@ class TensorizerAgent:
288
model_args.torch_dtype = self.tensorizer_config.dtype
289
assert self.tensorizer_config.model_class is not None
290
# TODO: Do we need to consider old-style model class?
291
- with no_init_or_tensor(), set_current_vllm_config(self.vllm_config):
291
+ with no_init_or_tensor(), set_current_vllm_config(self.vllm_config,
292
+ check_compile=True):
293
return self.tensorizer_config.model_class(
294
vllm_config=self.vllm_config, )
295
vllm/worker/worker_base.py
+13
-4
@@ -8,7 +8,8 @@ import cloudpickle
8
import torch
9
import torch.nn as nn
10
11
-from vllm.config import ObservabilityConfig, VllmConfig
11
+from vllm.config import (ObservabilityConfig, VllmConfig,
12
+ set_current_vllm_config)
13
from vllm.distributed import broadcast_tensor_dict, get_pp_group, get_tp_group
14
from vllm.logger import init_logger
15
from vllm.lora.request import LoRARequest
@@ -498,8 +499,11 @@ class WorkerWrapperBase:
499
group.
500
"""
501
self.rpc_rank = rpc_rank
501
- self.vllm_config = vllm_config
502
self.worker: Optional[WorkerBase] = None
503
+ # do not store this `vllm_config`, `init_worker` will set the final
504
+ # one. TODO: investigate if we can remove this field in
505
+ # `WorkerWrapperBase`, `init_cached_hf_modules` should be
506
+ # unnecessary now.
507
if vllm_config.model_config is not None:
508
# it can be None in tests
509
trust_remote_code = vllm_config.model_config.trust_remote_code
@@ -533,6 +537,9 @@ class WorkerWrapperBase:
537
Arguments are passed to the worker class constructor.
538
"""
539
kwargs = all_kwargs[self.rpc_rank]
540
+ self.vllm_config = kwargs.get("vllm_config", None)
541
+ assert self.vllm_config is not None, (
542
+ "vllm_config is required to initialize the worker")
543
enable_trace_function_call_for_thread(self.vllm_config)
544
545
from vllm.plugins import load_general_plugins
@@ -546,8 +553,10 @@ class WorkerWrapperBase:
553
bytes)
554
worker_class = cloudpickle.loads(
555
self.vllm_config.parallel_config.worker_cls)
549
- self.worker = worker_class(**kwargs)
550
- assert self.worker is not None
556
+ with set_current_vllm_config(self.vllm_config):
557
+ # To make vLLM config available during worker initialization
558
+ self.worker = worker_class(**kwargs)
559
+ assert self.worker is not None
560
561
def execute_method(self, method: Union[str, bytes], *args, **kwargs):
562
try: