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 14, 00:30 UTC
[Model] Introduce Kimi Linear to vLLM (#27809)
Signed-off-by: lizhiyuan <lizhiyuan@moonshot.cn> Signed-off-by: Zhiyuan Li <uniartisan2017@gmail.com>
Zhiyuan Li committed
Oct 30, 2025 at 21:02 UTC
4e68cc9b6aa2b9cfe8d799c2b1cd156a01bca438
15 files changed
+1326
-49
docs/models/supported_models.md
+1
@@ -382,6 +382,7 @@ th {
382
| `InternLM3ForCausalLM` | InternLM3 | `internlm/internlm3-8b-instruct`, etc. | ✅︎ | ✅︎ |
383
| `JAISLMHeadModel` | Jais | `inceptionai/jais-13b`, `inceptionai/jais-13b-chat`, `inceptionai/jais-30b-v3`, `inceptionai/jais-30b-chat-v3`, etc. | | ✅︎ |
384
| `JambaForCausalLM` | Jamba | `ai21labs/AI21-Jamba-1.5-Large`, `ai21labs/AI21-Jamba-1.5-Mini`, `ai21labs/Jamba-v0.1`, etc. | ✅︎ | ✅︎ |
385
+| `KimiLinearForCausalLM` | Kimi-Linear-48B-A3B-Base, Kimi-Linear-48B-A3B-Instruct | `moonshotai/Kimi-Linear-48B-A3B-Base`, `moonshotai/Kimi-Linear-48B-A3B-Instruct` | | ✅︎ |
386
| `Lfm2ForCausalLM` | LFM2 | `LiquidAI/LFM2-1.2B`, `LiquidAI/LFM2-700M`, `LiquidAI/LFM2-350M`, etc. | ✅︎ | ✅︎ |
387
| `Lfm2MoeForCausalLM` | LFM2MoE | `LiquidAI/LFM2-8B-A1B-preview`, etc. | ✅︎ | ✅︎ |
388
| `LlamaForCausalLM` | Llama 3.1, Llama 3, Llama 2, LLaMA, Yi | `meta-llama/Meta-Llama-3.1-405B-Instruct`, `meta-llama/Meta-Llama-3.1-70B`, `meta-llama/Meta-Llama-3-70B-Instruct`, `meta-llama/Llama-2-70b-hf`, `01-ai/Yi-34B`, etc. | ✅︎ | ✅︎ |
tests/models/registry.py
+3
@@ -296,6 +296,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
296
"random": "ai21labs/Jamba-tiny-random",
297
},
298
),
299
+ "KimiLinearForCausalLM": _HfExamplesInfo(
300
+ "moonshotai/Kimi-Linear-48B-A3B-Instruct", trust_remote_code=True
301
+ ),
302
"Lfm2ForCausalLM": _HfExamplesInfo("LiquidAI/LFM2-1.2B"),
303
"Lfm2MoeForCausalLM": _HfExamplesInfo(
304
"LiquidAI/LFM2-8B-A1B", min_transformers_version="4.58"
vllm/config/compilation.py
+1
@@ -453,6 +453,7 @@ class CompilationConfig:
453
"vllm::linear_attention",
454
"vllm::plamo2_mamba_mixer",
455
"vllm::gdn_attention",
456
+ "vllm::kda_attention",
457
"vllm::sparse_attn_indexer",
458
]
459
vllm/config/model.py
+1
@@ -1236,6 +1236,7 @@ class ModelConfig:
1236
"deepseek_v32",
1237
"deepseek_mtp",
1238
"kimi_k2",
1239
+ "kimi_linear",
1240
"longcat_flash",
1241
):
1242
return self.hf_text_config.kv_lora_rank is not None
vllm/model_executor/layers/fla/ops/kda.py
+1
-1
@@ -1304,7 +1304,7 @@ def kda_gate_fwd_kernel(
1304
tl.store(y_ptr, b_y.to(y.dtype.element_ty), boundary_check=(0, 1))
1305
1306
1307
-def kda_gate_fwd(
1307
+def fused_kda_gate(
1308
g: torch.Tensor,
1309
A: torch.Tensor,
1310
head_k_dim: int,
vllm/model_executor/layers/kda.py
new
+426
@@ -0,0 +1,426 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+
4
+import torch
5
+from einops import rearrange
6
+from torch import nn
7
+
8
+from vllm.attention import AttentionBackend
9
+from vllm.attention.backends.abstract import AttentionMetadata
10
+from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
11
+from vllm.distributed import (
12
+ divide,
13
+ get_tensor_model_parallel_rank,
14
+ get_tensor_model_parallel_world_size,
15
+)
16
+from vllm.forward_context import ForwardContext, get_forward_context
17
+from vllm.logger import init_logger
18
+from vllm.model_executor.model_loader.weight_utils import sharded_weight_loader
19
+from vllm.model_executor.utils import set_weight_attrs
20
+from vllm.utils.torch_utils import direct_register_custom_op
21
+from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
22
+
23
+from .fla.ops.kda import (
24
+ FusedRMSNormGated,
25
+ chunk_kda,
26
+ fused_kda_gate,
27
+ fused_recurrent_kda,
28
+)
29
+from .linear import (
30
+ ColumnParallelLinear,
31
+ ReplicatedLinear,
32
+ RowParallelLinear,
33
+)
34
+from .mamba.abstract import MambaBase
35
+from .mamba.mamba_utils import MambaStateDtypeCalculator, MambaStateShapeCalculator
36
+from .mamba.ops.causal_conv1d import causal_conv1d_fn, causal_conv1d_update
37
+from .quantization.base_config import QuantizationConfig
38
+
39
+logger = init_logger(__name__)
40
+
41
+
42
+def kda_attention(
43
+ hidden_states: torch.Tensor,
44
+ output: torch.Tensor,
45
+ layer_name: str,
46
+) -> None:
47
+ forward_context: ForwardContext = get_forward_context()
48
+ self = forward_context.no_compile_layers[layer_name]
49
+ self._forward(hidden_states=hidden_states, output=output)
50
+
51
+
52
+def kda_attention_fake(
53
+ hidden_states: torch.Tensor,
54
+ output: torch.Tensor,
55
+ layer_name: str,
56
+) -> None:
57
+ return
58
+
59
+
60
+direct_register_custom_op(
61
+ op_name="kda_attention",
62
+ op_func=kda_attention,
63
+ mutates_args=["output"],
64
+ fake_impl=kda_attention_fake,
65
+)
66
+
67
+
68
+class KimiDeltaAttention(nn.Module, MambaBase):
69
+ @property
70
+ def mamba_type(self) -> str:
71
+ return "linear_attention"
72
+
73
+ def get_attn_backend(self) -> type["AttentionBackend"]:
74
+ from vllm.v1.attention.backends.gdn_attn import GDNAttentionBackend
75
+
76
+ return GDNAttentionBackend
77
+
78
+ def get_state_dtype(
79
+ self,
80
+ ) -> tuple[torch.dtype, torch.dtype, torch.dtype, torch.dtype]:
81
+ if self.model_config is None or self.cache_config is None:
82
+ raise ValueError("model_config and cache_config must be set")
83
+ return MambaStateDtypeCalculator.kda_state_dtype(
84
+ self.model_config.dtype, self.cache_config.mamba_cache_dtype
85
+ )
86
+
87
+ def get_state_shape(
88
+ self,
89
+ ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
90
+ return MambaStateShapeCalculator.kda_state_shape(
91
+ self.tp_size, self.num_heads, self.head_dim, conv_kernel_size=self.conv_size
92
+ )
93
+
94
+ def __init__(
95
+ self,
96
+ layer_idx: int,
97
+ hidden_size: int,
98
+ quant_config: QuantizationConfig | None = None,
99
+ cache_config: CacheConfig | None = None,
100
+ model_config: ModelConfig | None = None,
101
+ rms_norm_eps: float = 1e-5,
102
+ prefix: str = "",
103
+ **kwargs,
104
+ ) -> None:
105
+ super().__init__()
106
+ self.tp_size = get_tensor_model_parallel_world_size()
107
+ self.tp_rank = get_tensor_model_parallel_rank()
108
+ self.hidden_size = hidden_size
109
+ self.model_config = model_config
110
+ self.cache_config = cache_config
111
+ if model_config is None:
112
+ raise ValueError("model_config must be provided")
113
+ kda_config = model_config.linear_attn_config
114
+ self.head_dim = kda_config["head_dim"]
115
+ self.num_heads = kda_config["num_heads"]
116
+ self.layer_idx = layer_idx
117
+ self.prefix = prefix
118
+ assert self.num_heads % self.tp_size == 0
119
+ self.local_num_heads = divide(self.num_heads, self.tp_size)
120
+
121
+ projection_size = self.head_dim * self.num_heads
122
+ self.conv_size = kda_config["short_conv_kernel_size"]
123
+
124
+ self.q_proj = ColumnParallelLinear(
125
+ self.hidden_size,
126
+ projection_size,
127
+ bias=False,
128
+ quant_config=quant_config,
129
+ prefix=f"{prefix}.q_proj",
130
+ )
131
+ self.k_proj = ColumnParallelLinear(
132
+ self.hidden_size,
133
+ projection_size,
134
+ bias=False,
135
+ quant_config=quant_config,
136
+ prefix=f"{prefix}.k_proj",
137
+ )
138
+ self.v_proj = ColumnParallelLinear(
139
+ self.hidden_size,
140
+ projection_size,
141
+ bias=False,
142
+ quant_config=quant_config,
143
+ prefix=f"{prefix}.v_proj",
144
+ )
145
+
146
+ self.f_a_proj = ReplicatedLinear(
147
+ self.hidden_size,
148
+ self.head_dim,
149
+ bias=False,
150
+ quant_config=quant_config,
151
+ prefix=f"{prefix}.f_a_proj",
152
+ )
153
+
154
+ self.f_b_proj = ColumnParallelLinear(
155
+ self.head_dim,
156
+ projection_size,
157
+ bias=False,
158
+ quant_config=quant_config,
159
+ prefix=f"{prefix}.f_b_proj",
160
+ )
161
+ self.dt_bias = nn.Parameter(
162
+ torch.empty(divide(projection_size, self.tp_size), dtype=torch.float32)
163
+ )
164
+
165
+ set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
166
+
167
+ self.b_proj = ColumnParallelLinear(
168
+ self.hidden_size,
169
+ self.num_heads,
170
+ bias=False,
171
+ quant_config=quant_config,
172
+ prefix=f"{prefix}.b_proj",
173
+ )
174
+
175
+ self.q_conv1d = ColumnParallelLinear(
176
+ input_size=self.conv_size,
177
+ output_size=projection_size,
178
+ bias=False,
179
+ params_dtype=torch.float32,
180
+ prefix=f"{prefix}.q_conv1d",
181
+ )
182
+ self.k_conv1d = ColumnParallelLinear(
183
+ input_size=self.conv_size,
184
+ output_size=projection_size,
185
+ bias=False,
186
+ params_dtype=torch.float32,
187
+ prefix=f"{prefix}.k_conv1d",
188
+ )
189
+ self.v_conv1d = ColumnParallelLinear(
190
+ input_size=self.conv_size,
191
+ output_size=projection_size,
192
+ bias=False,
193
+ params_dtype=torch.float32,
194
+ prefix=f"{prefix}.v_conv1d",
195
+ )
196
+ # unsqueeze to fit conv1d weights shape into the linear weights shape.
197
+ # Can't do this in `weight_loader` since it already exists in
198
+ # `ColumnParallelLinear` and `set_weight_attrs`
199
+ # doesn't allow to override it
200
+ self.q_conv1d.weight.data = self.q_conv1d.weight.data.unsqueeze(1)
201
+ self.k_conv1d.weight.data = self.k_conv1d.weight.data.unsqueeze(1)
202
+ self.v_conv1d.weight.data = self.v_conv1d.weight.data.unsqueeze(1)
203
+
204
+ self.A_log = nn.Parameter(
205
+ torch.empty(1, 1, self.local_num_heads, 1, dtype=torch.float32)
206
+ )
207
+ set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(2)})
208
+
209
+ self.g_a_proj = ReplicatedLinear(
210
+ self.hidden_size,
211
+ self.head_dim,
212
+ bias=False,
213
+ quant_config=quant_config,
214
+ prefix=f"{prefix}.g_a_proj",
215
+ )
216
+ self.g_b_proj = ColumnParallelLinear(
217
+ self.head_dim,
218
+ projection_size,
219
+ bias=False,
220
+ quant_config=quant_config,
221
+ prefix=f"{prefix}.g_b_proj",
222
+ )
223
+ self.o_norm = FusedRMSNormGated(
224
+ self.head_dim, eps=rms_norm_eps, activation="sigmoid"
225
+ )
226
+ self.o_proj = RowParallelLinear(
227
+ projection_size,
228
+ self.hidden_size,
229
+ bias=False,
230
+ quant_config=quant_config,
231
+ prefix=f"{prefix}.o_proj",
232
+ )
233
+
234
+ compilation_config = get_current_vllm_config().compilation_config
235
+ if prefix in compilation_config.static_forward_context:
236
+ raise ValueError(f"Duplicate layer name: {prefix}")
237
+ compilation_config.static_forward_context[prefix] = self
238
+
239
+ def forward(
240
+ self,
241
+ hidden_states: torch.Tensor,
242
+ positions: torch.Tensor,
243
+ output: torch.Tensor,
244
+ ) -> None:
245
+ return torch.ops.vllm.kda_attention(
246
+ hidden_states,
247
+ output,
248
+ self.prefix,
249
+ )
250
+
251
+ def _forward(
252
+ self,
253
+ hidden_states: torch.Tensor,
254
+ output: torch.Tensor,
255
+ ) -> None:
256
+ forward_context = get_forward_context()
257
+ attn_metadata: AttentionMetadata = forward_context.attn_metadata
258
+
259
+ if attn_metadata is None:
260
+ # V1 profile run
261
+ # Mimic the memory allocation in the real run
262
+ q = torch.empty_like(hidden_states)
263
+ k = torch.empty_like(hidden_states)
264
+ v = torch.empty_like(hidden_states)
265
+ g = hidden_states.new_empty(
266
+ hidden_states.size(0),
267
+ self.local_num_heads,
268
+ self.head_dim,
269
+ dtype=torch.float32,
270
+ )
271
+ beta = torch.empty(
272
+ hidden_states.size(0), self.local_num_heads, dtype=torch.float32
273
+ )
274
+ core_attn_out = torch.empty_like(hidden_states)
275
+ return
276
+
277
+ assert isinstance(attn_metadata, dict)
278
+ attn_metadata = attn_metadata[self.prefix]
279
+ assert isinstance(attn_metadata, GDNAttentionMetadata)
280
+ has_initial_state = attn_metadata.has_initial_state
281
+ non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
282
+ non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor # noqa: E501
283
+ constant_caches = self.kv_cache[forward_context.virtual_engine]
284
+
285
+ (conv_state_q, conv_state_k, conv_state_v, recurrent_state) = constant_caches
286
+ # deal with strides
287
+ conv_state_q = conv_state_q.transpose(-1, -2)
288
+ conv_state_k = conv_state_k.transpose(-1, -2)
289
+ conv_state_v = conv_state_v.transpose(-1, -2)
290
+
291
+ q_proj_states = self.q_proj(hidden_states)[0]
292
+ k_proj_states = self.k_proj(hidden_states)[0]
293
+ v_proj_states = self.v_proj(hidden_states)[0]
294
+
295
+ q_conv_weights = self.q_conv1d.weight.view(
296
+ self.q_conv1d.weight.size(0), self.q_conv1d.weight.size(2)
297
+ )
298
+ k_conv_weights = self.k_conv1d.weight.view(
299
+ self.k_conv1d.weight.size(0), self.k_conv1d.weight.size(2)
300
+ )
301
+ v_conv_weights = self.v_conv1d.weight.view(
302
+ self.v_conv1d.weight.size(0), self.v_conv1d.weight.size(2)
303
+ )
304
+ if attn_metadata.num_prefills > 0:
305
+ q_proj_states = q_proj_states.transpose(0, 1)
306
+ k_proj_states = k_proj_states.transpose(0, 1)
307
+ v_proj_states = v_proj_states.transpose(0, 1)
308
+ q = causal_conv1d_fn(
309
+ q_proj_states,
310
+ q_conv_weights,
311
+ self.q_conv1d.bias,
312
+ activation="silu",
313
+ conv_states=conv_state_q,
314
+ has_initial_state=has_initial_state,
315
+ cache_indices=non_spec_state_indices_tensor,
316
+ query_start_loc=non_spec_query_start_loc,
317
+ metadata=attn_metadata,
318
+ ).transpose(0, 1)
319
+ k = causal_conv1d_fn(
320
+ k_proj_states,
321
+ k_conv_weights,
322
+ self.k_conv1d.bias,
323
+ activation="silu",
324
+ conv_states=conv_state_k,
325
+ has_initial_state=has_initial_state,
326
+ cache_indices=non_spec_state_indices_tensor,
327
+ query_start_loc=non_spec_query_start_loc,
328
+ metadata=attn_metadata,
329
+ ).transpose(0, 1)
330
+ v = causal_conv1d_fn(
331
+ v_proj_states,
332
+ v_conv_weights,
333
+ self.v_conv1d.bias,
334
+ activation="silu",
335
+ conv_states=conv_state_v,
336
+ has_initial_state=has_initial_state,
337
+ cache_indices=non_spec_state_indices_tensor,
338
+ query_start_loc=non_spec_query_start_loc,
339
+ metadata=attn_metadata,
340
+ ).transpose(0, 1)
341
+ else:
342
+ decode_conv_indices = non_spec_state_indices_tensor[
343
+ : attn_metadata.num_decodes
344
+ ]
345
+ q = causal_conv1d_update(
346
+ q_proj_states,
347
+ conv_state_q,
348
+ q_conv_weights,
349
+ self.q_conv1d.bias,
350
+ activation="silu",
351
+ conv_state_indices=decode_conv_indices,
352
+ validate_data=True,
353
+ )
354
+ k = causal_conv1d_update(
355
+ k_proj_states,
356
+ conv_state_k,
357
+ k_conv_weights,
358
+ self.k_conv1d.bias,
359
+ activation="silu",
360
+ conv_state_indices=decode_conv_indices,
361
+ validate_data=True,
362
+ )
363
+ v = causal_conv1d_update(
364
+ v_proj_states,
365
+ conv_state_v,
366
+ v_conv_weights,
367
+ self.v_conv1d.bias,
368
+ activation="silu",
369
+ conv_state_indices=decode_conv_indices,
370
+ validate_data=True,
371
+ )
372
+
373
+ q, k, v = map(
374
+ lambda x: rearrange(x, "n (h d) -> 1 n h d", d=self.head_dim), (q, k, v)
375
+ )
376
+
377
+ beta = self.b_proj(hidden_states)[0].float().sigmoid()
378
+
379
+ g = self.f_b_proj(self.f_a_proj(hidden_states)[0])[0]
380
+ g = fused_kda_gate(g, self.A_log, self.head_dim, g_bias=self.dt_bias)
381
+
382
+ beta = beta.unsqueeze(0)
383
+ g = g.unsqueeze(0)
384
+
385
+ if attn_metadata.num_prefills > 0:
386
+ zero_idx = non_spec_state_indices_tensor[~has_initial_state]
387
+ recurrent_state[zero_idx] = 0
388
+ initial_state = recurrent_state[non_spec_state_indices_tensor].contiguous()
389
+ (
390
+ core_attn_out_non_spec,
391
+ last_recurrent_state,
392
+ ) = chunk_kda(
393
+ q=q,
394
+ k=k,
395
+ v=v,
396
+ g=g,
397
+ beta=beta,
398
+ initial_state=initial_state,
399
+ output_final_state=True,
400
+ use_qk_l2norm_in_kernel=True,
401
+ cu_seqlens=non_spec_query_start_loc,
402
+ )
403
+ # Init cache
404
+ recurrent_state[non_spec_state_indices_tensor] = last_recurrent_state
405
+ else:
406
+ (
407
+ core_attn_out_non_spec,
408
+ last_recurrent_state,
409
+ ) = fused_recurrent_kda(
410
+ q=q,
411
+ k=k,
412
+ v=v,
413
+ g=g,
414
+ beta=beta,
415
+ initial_state=recurrent_state,
416
+ use_qk_l2norm_in_kernel=True,
417
+ cu_seqlens=non_spec_query_start_loc,
418
+ ssm_state_indices=non_spec_state_indices_tensor,
419
+ )
420
+
421
+ g_proj_states = self.g_b_proj(self.g_a_proj(hidden_states)[0])[0]
422
+ g = rearrange(g_proj_states, "... (h d) -> ... h d", d=self.head_dim)
423
+ core_attn_out = self.o_norm(core_attn_out_non_spec, g)
424
+ core_attn_out = rearrange(core_attn_out, "1 n h d -> n (h d)")
425
+
426
+ output[:] = self.o_proj(core_attn_out)[0]
vllm/model_executor/layers/mamba/mamba_utils.py
+41
@@ -80,6 +80,15 @@ class MambaStateDtypeCalculator:
80
state_dtype = get_kv_cache_torch_dtype(mamba_cache_dtype, model_dtype)
81
return (state_dtype, state_dtype)
82
83
+ @classmethod
84
+ def kda_state_dtype(
85
+ cls,
86
+ model_dtype: ModelDType | torch.dtype,
87
+ mamba_cache_dtype: MambaDType,
88
+ ):
89
+ state_dtype = get_kv_cache_torch_dtype(mamba_cache_dtype, model_dtype)
90
+ return (state_dtype, state_dtype, state_dtype, torch.float32)
91
+
92
93
class MambaStateShapeCalculator:
94
@classmethod
@@ -182,3 +191,35 @@ class MambaStateShapeCalculator:
191
head_v_dim,
192
)
193
return conv_state_shape, temporal_state_shape
194
+
195
+ @classmethod
196
+ def kda_state_shape(
197
+ cls,
198
+ tp_world_size: int,
199
+ num_heads: int,
200
+ head_dim: int,
201
+ num_k_heads: int | None = None,
202
+ head_k_dim: int | None = None,
203
+ conv_kernel_size: int = 4,
204
+ num_spec: int = 0,
205
+ ) -> tuple[tuple[int, int], tuple[int, int], tuple[int, int], tuple[int, int, int]]:
206
+ if num_k_heads is None:
207
+ num_k_heads = num_heads
208
+ if head_k_dim is None:
209
+ head_k_dim = head_dim
210
+
211
+ proj_size = num_heads * head_dim
212
+ proj_k_size = num_k_heads * head_k_dim
213
+
214
+ conv_state_shape = (divide(proj_size, tp_world_size), conv_kernel_size - 1)
215
+ conv_state_k_shape = (divide(proj_k_size, tp_world_size), conv_kernel_size - 1)
216
+ recurrent_state_shape = (divide(num_heads, tp_world_size), head_dim, head_dim)
217
+
218
+ conv_state_shape = conv_state_shape[1], conv_state_shape[0]
219
+ conv_state_k_shape = conv_state_k_shape[1], conv_state_k_shape[0]
220
+ return (
221
+ conv_state_shape,
222
+ conv_state_k_shape,
223
+ conv_state_k_shape,
224
+ recurrent_state_shape,
225
+ )
vllm/model_executor/layers/mla.py
+4
-3
@@ -147,9 +147,10 @@ class MultiHeadLatentAttentionWrapper(CustomOp):
147
# Add head dim of 1 to k_pe
148
k_pe = k_pe.unsqueeze(1)
149
150
- q[..., self.qk_nope_head_dim :], k_pe = self.rotary_emb(
151
- positions, q[..., self.qk_nope_head_dim :], k_pe
152
- )
150
+ if self.rotary_emb is not None:
151
+ q[..., self.qk_nope_head_dim :], k_pe = self.rotary_emb(
152
+ positions, q[..., self.qk_nope_head_dim :], k_pe
153
+ )
154
155
if self.indexer and self.is_sparse:
156
_topk_indices = self.indexer(hidden_states, q_c, positions, self.rotary_emb)
vllm/model_executor/models/config.py
+25
-26
@@ -1,6 +1,7 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
from copy import deepcopy
4
+from math import lcm
5
from typing import TYPE_CHECKING
6
7
import vllm.envs as envs
@@ -8,7 +9,7 @@ from vllm.logger import init_logger
9
from vllm.model_executor.models import ModelRegistry
10
from vllm.utils.math_utils import cdiv, round_up
11
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
11
-from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec
12
+from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec, MLAAttentionSpec
13
14
if TYPE_CHECKING:
15
from vllm.config import VllmConfig
@@ -347,12 +348,28 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
348
kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
349
350
# get attention page size (for 1 token)
350
- attn_page_size_1_token = FullAttentionSpec(
351
- block_size=1,
352
- num_kv_heads=model_config.get_num_kv_heads(parallel_config),
353
- head_size=model_config.get_head_size(),
354
- dtype=kv_cache_dtype,
355
- ).page_size_bytes
351
+ # Attention backend constraints:
352
+ # - FlashAttention (FA) requires block size to be multiple of 16
353
+ # - MLA (Multi-head Latent Attention) requires larger alignment:
354
+ # * CUTLASS_MLA backend: kernel_block_size 128 alignment
355
+ # * Other MLA backends: kernel_block_size 64 alignment
356
+ if model_config.use_mla:
357
+ use_cutlass_mla = envs.VLLM_ATTENTION_BACKEND == "CUTLASS_MLA"
358
+ kernel_block_alignment_size = 128 if use_cutlass_mla else 64
359
+ attn_page_size_1_token = MLAAttentionSpec(
360
+ block_size=1,
361
+ num_kv_heads=model_config.get_num_kv_heads(parallel_config),
362
+ head_size=model_config.get_head_size(),
363
+ dtype=kv_cache_dtype,
364
+ ).page_size_bytes
365
+ else:
366
+ kernel_block_alignment_size = 16
367
+ attn_page_size_1_token = FullAttentionSpec(
368
+ block_size=1,
369
+ num_kv_heads=model_config.get_num_kv_heads(parallel_config),
370
+ head_size=model_config.get_head_size(),
371
+ dtype=kv_cache_dtype,
372
+ ).page_size_bytes
373
374
model_cls, _ = ModelRegistry.resolve_model_cls(
375
model_config.architecture,
@@ -372,17 +389,6 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
389
if mamba_page_size == 0:
390
return
391
375
- # Attention backend constraints:
376
- # - FlashAttention (FA) requires block size to be multiple of 16
377
- # - MLA (Multi-head Latent Attention) requires larger alignment:
378
- # * CUTLASS_MLA backend: 128-byte alignment
379
- # * Other MLA backends: 64-byte alignment
380
- if model_config.use_mla:
381
- use_cutlass_mla = envs.VLLM_ATTENTION_BACKEND == "CUTLASS_MLA"
382
- kernel_block_alignment_size = 128 if use_cutlass_mla else 64
383
- else:
384
- kernel_block_alignment_size = 16
385
-
392
if cache_config.enable_prefix_caching:
393
# With prefix caching, select attention block size to
394
# optimize for mamba kernel performance
@@ -400,15 +406,8 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
406
# easily by changing the way we layout chunks in the
407
# mamba2 kernels.
408
403
- from math import gcd
404
-
405
- def lcm(a, b):
406
- return a * b // gcd(a, b)
407
-
408
- base_chunk_size = mamba_block_size or model_config.get_mamba_chunk_size()
409
-
409
+ base_chunk_size = model_config.get_mamba_chunk_size()
410
attn_tokens_per_mamba_state = cdiv(mamba_page_size, attn_page_size_1_token)
411
-
411
chunk_size = lcm(base_chunk_size, kernel_block_alignment_size)
412
attn_block_size = chunk_size * cdiv(attn_tokens_per_mamba_state, chunk_size)
413
cache_config.mamba_block_size = attn_block_size
vllm/model_executor/models/kimi_linear.py
new
+663
@@ -0,0 +1,663 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+
4
+from collections.abc import Iterable
5
+from typing import Any
6
+
7
+import torch
8
+from torch import nn
9
+
10
+from vllm.compilation.decorators import support_torch_compile
11
+from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
12
+from vllm.distributed import (
13
+ get_pp_group,
14
+ get_tensor_model_parallel_world_size,
15
+ tensor_model_parallel_all_reduce,
16
+)
17
+from vllm.logger import init_logger
18
+from vllm.model_executor.layers.activation import SiluAndMul
19
+from vllm.model_executor.layers.fused_moe import FusedMoE
20
+from vllm.model_executor.layers.kda import KimiDeltaAttention
21
+from vllm.model_executor.layers.layernorm import RMSNorm
22
+from vllm.model_executor.layers.linear import (
23
+ ColumnParallelLinear,
24
+ MergedColumnParallelLinear,
25
+ QKVParallelLinear,
26
+ ReplicatedLinear,
27
+ RowParallelLinear,
28
+)
29
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
30
+from vllm.model_executor.layers.mamba.mamba_utils import (
31
+ MambaStateDtypeCalculator,
32
+ MambaStateShapeCalculator,
33
+)
34
+from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper
35
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
36
+from vllm.model_executor.layers.vocab_parallel_embedding import (
37
+ ParallelLMHead,
38
+ VocabParallelEmbedding,
39
+)
40
+from vllm.model_executor.model_loader.weight_utils import (
41
+ default_weight_loader,
42
+ maybe_remap_kv_scale_name,
43
+)
44
+from vllm.sequence import IntermediateTensors
45
+from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig
46
+
47
+from .interfaces import HasInnerState, IsHybrid, MixtureOfExperts, SupportsPP
48
+from .utils import (
49
+ PPMissingLayer,
50
+ is_pp_missing_parameter,
51
+ make_layers,
52
+ maybe_prefix,
53
+)
54
+
55
+logger = init_logger(__name__)
56
+
57
+
58
+class KimiMLP(nn.Module):
59
+ def __init__(
60
+ self,
61
+ hidden_size: int,
62
+ intermediate_size: int,
63
+ hidden_act: str,
64
+ quant_config: QKVParallelLinear | None = None,
65
+ reduce_results: bool = True,
66
+ prefix: str = "",
67
+ ) -> None:
68
+ super().__init__()
69
+
70
+ self.gate_up_proj = MergedColumnParallelLinear(
71
+ hidden_size,
72
+ [intermediate_size] * 2,
73
+ bias=False,
74
+ quant_config=quant_config,
75
+ prefix=f"{prefix}.gate_up_proj",
76
+ )
77
+ self.down_proj = RowParallelLinear(
78
+ intermediate_size,
79
+ hidden_size,
80
+ bias=False,
81
+ quant_config=quant_config,
82
+ reduce_results=reduce_results,
83
+ prefix=f"{prefix}.down_proj",
84
+ )
85
+ if hidden_act != "silu":
86
+ raise ValueError(
87
+ f"Unsupported activation: {hidden_act}. Only silu is supported for now."
88
+ )
89
+ self.act_fn = SiluAndMul()
90
+
91
+ def forward(self, x):
92
+ gate_up, _ = self.gate_up_proj(x)
93
+ x = self.act_fn(gate_up)
94
+ x, _ = self.down_proj(x)
95
+ return x
96
+
97
+
98
+class KimiMoE(nn.Module):
99
+ def __init__(
100
+ self,
101
+ config: KimiLinearConfig,
102
+ quant_config: QuantizationConfig | None = None,
103
+ prefix: str = "",
104
+ layer_idx: int = 0,
105
+ ):
106
+ super().__init__()
107
+ hidden_size = config.hidden_size
108
+ intermediate_size = config.intermediate_size
109
+ moe_intermediate_size = config.moe_intermediate_size
110
+ num_experts = config.num_experts
111
+ moe_renormalize = config.moe_renormalize
112
+ self.tp_size = get_tensor_model_parallel_world_size()
113
+ self.routed_scaling_factor = config.routed_scaling_factor
114
+ self.num_shared_experts = config.num_shared_experts
115
+ self.layer_idx = layer_idx
116
+
117
+ if config.hidden_act != "silu":
118
+ raise ValueError(
119
+ f"Unsupported activation: {config.hidden_act}. "
120
+ "Only silu is supported for now."
121
+ )
122
+
123
+ # Gate always runs at half / full precision for now.
124
+ self.gate = ReplicatedLinear(
125
+ hidden_size,
126
+ num_experts,
127
+ bias=False,
128
+ quant_config=None,
129
+ prefix=f"{prefix}.gate",
130
+ )
131
+
132
+ self.gate.e_score_correction_bias = nn.Parameter(torch.empty(num_experts))
133
+
134
+ self.experts = FusedMoE(
135
+ num_experts=num_experts,
136
+ top_k=config.num_experts_per_token,
137
+ hidden_size=hidden_size,
138
+ intermediate_size=moe_intermediate_size,
139
+ reduce_results=False,
140
+ renormalize=moe_renormalize,
141
+ quant_config=quant_config,
142
+ use_grouped_topk=config.use_grouped_topk,
143
+ num_expert_group=config.num_expert_group,
144
+ topk_group=config.topk_group,
145
+ prefix=f"{prefix}.experts",
146
+ scoring_func=config.moe_router_activation_func,
147
+ e_score_correction_bias=self.gate.e_score_correction_bias,
148
+ )
149
+
150
+ if self.num_shared_experts is not None:
151
+ intermediate_size = moe_intermediate_size * self.num_shared_experts
152
+ self.shared_experts = KimiMLP(
153
+ hidden_size=config.hidden_size,
154
+ intermediate_size=intermediate_size,
155
+ hidden_act=config.hidden_act,
156
+ quant_config=quant_config,
157
+ reduce_results=False,
158
+ )
159
+
160
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
161
+ num_tokens, hidden_size = hidden_states.shape
162
+ hidden_states = hidden_states.view(-1, hidden_size)
163
+ if self.num_shared_experts is not None:
164
+ shared_output = self.shared_experts(hidden_states)
165
+ router_logits, _ = self.gate(hidden_states)
166
+ final_hidden_states = (
167
+ self.experts(hidden_states=hidden_states, router_logits=router_logits)
168
+ * self.routed_scaling_factor
169
+ )
170
+ if shared_output is not None:
171
+ final_hidden_states = final_hidden_states + shared_output
172
+
173
+ if self.tp_size > 1:
174
+ final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
175
+ return final_hidden_states.view(num_tokens, hidden_size)
176
+
177
+
178
+class KimiMLAAttention(nn.Module):
179
+ """
180
+ Main reference: DeepseekV2 vllm Implementation
181
+ """
182
+
183
+ def __init__(
184
+ self,
185
+ config: KimiLinearConfig,
186
+ hidden_size: int,
187
+ num_heads: int,
188
+ qk_nope_head_dim: int,
189
+ qk_rope_head_dim: int,
190
+ v_head_dim: int,
191
+ q_lora_rank: int | None,
192
+ kv_lora_rank: int,
193
+ rope_theta: float = 10000,
194
+ use_nope: bool = False,
195
+ rope_scaling: dict[str, Any] | None = None,
196
+ cache_config: CacheConfig | None = None,
197
+ quant_config: QuantizationConfig | None = None,
198
+ prefix: str = "",
199
+ **kwargs,
200
+ ) -> None:
201
+ super().__init__()
202
+ self.hidden_size = hidden_size
203
+ self.qk_nope_head_dim = qk_nope_head_dim
204
+ self.qk_rope_head_dim = qk_rope_head_dim
205
+ self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
206
+ self.v_head_dim = v_head_dim
207
+ self.q_lora_rank = q_lora_rank
208
+ self.kv_lora_rank = kv_lora_rank
209
+ self.num_heads = num_heads
210
+ tp_size = get_tensor_model_parallel_world_size()
211
+ self.num_local_heads = num_heads // tp_size
212
+ self.scaling = self.qk_head_dim**-0.5
213
+ self.rope_theta = rope_theta
214
+ self.use_nope = use_nope
215
+ assert self.use_nope is True
216
+ assert self.q_lora_rank is None
217
+ assert rope_scaling is None
218
+ assert num_heads % tp_size == 0
219
+ self.kv_a_proj_with_mqa = ReplicatedLinear(
220
+ self.hidden_size,
221
+ self.kv_lora_rank + self.qk_rope_head_dim,
222
+ bias=False,
223
+ quant_config=quant_config,
224
+ prefix=f"{prefix}.kv_a_proj_with_mqa",
225
+ )
226
+ self.q_proj = ColumnParallelLinear(
227
+ self.hidden_size,
228
+ self.num_heads * self.qk_head_dim,
229
+ bias=False,
230
+ quant_config=quant_config,
231
+ prefix=f"{prefix}.q_proj",
232
+ )
233
+ self.kv_a_layernorm = RMSNorm(
234
+ self.kv_lora_rank,
235
+ eps=config.rms_norm_eps,
236
+ )
237
+ self.kv_b_proj = ColumnParallelLinear(
238
+ self.kv_lora_rank,
239
+ self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
240
+ bias=False,
241
+ quant_config=quant_config,
242
+ prefix=f"{prefix}.kv_b_proj",
243
+ )
244
+ self.o_proj = RowParallelLinear(
245
+ self.num_heads * self.v_head_dim,
246
+ self.hidden_size,
247
+ bias=False,
248
+ quant_config=quant_config,
249
+ prefix=f"{prefix}.o_proj",
250
+ )
251
+
252
+ mla_modules = MLAModules(
253
+ kv_a_layernorm=self.kv_a_layernorm,
254
+ kv_b_proj=self.kv_b_proj,
255
+ rotary_emb=None,
256
+ o_proj=self.o_proj,
257
+ fused_qkv_a_proj=None,
258
+ kv_a_proj_with_mqa=self.kv_a_proj_with_mqa,
259
+ q_a_layernorm=None,
260
+ q_b_proj=None,
261
+ q_proj=self.q_proj,
262
+ indexer=None,
263
+ is_sparse=False,
264
+ topk_indices_buffer=None,
265
+ )
266
+ self.mla_attn = MultiHeadLatentAttentionWrapper(
267
+ self.hidden_size,
268
+ self.num_local_heads,
269
+ self.scaling,
270
+ self.qk_nope_head_dim,
271
+ self.qk_rope_head_dim,
272
+ self.v_head_dim,
273
+ self.q_lora_rank,
274
+ self.kv_lora_rank,
275
+ mla_modules,
276
+ cache_config,
277
+ quant_config,
278
+ prefix,
279
+ )
280
+
281
+ def forward(
282
+ self,
283
+ positions: torch.Tensor,
284
+ hidden_states: torch.Tensor,
285
+ output: torch.Tensor,
286
+ ) -> None:
287
+ output[:] = self.mla_attn(positions, hidden_states)
288
+
289
+
290
+class KimiDecoderLayer(nn.Module):
291
+ def __init__(
292
+ self,
293
+ config: KimiLinearConfig,
294
+ layer_idx: int,
295
+ cache_config: CacheConfig | None = None,
296
+ quant_config: QuantizationConfig | None = None,
297
+ parallel_config: ParallelConfig | None = None,
298
+ model_config: ModelConfig | None = None,
299
+ prefix: str = "",
300
+ **kwargs,
301
+ ) -> None:
302
+ super().__init__()
303
+ self.hidden_size = config.hidden_size
304
+
305
+ self.is_moe = config.is_moe
306
+
307
+ if config.is_kda_layer(layer_idx):
308
+ self.self_attn = KimiDeltaAttention(
309
+ layer_idx=layer_idx,
310
+ hidden_size=config.hidden_size,
311
+ quant_config=quant_config,
312
+ cache_config=cache_config,
313
+ model_config=config,
314
+ prefix=f"{prefix}.self_attn",
315
+ )
316
+ else:
317
+ self.self_attn = KimiMLAAttention(
318
+ layer_idx=layer_idx,
319
+ hidden_size=self.hidden_size,
320
+ num_heads=config.num_attention_heads,
321
+ quant_config=quant_config,
322
+ cache_config=cache_config,
323
+ model_config=model_config,
324
+ prefix=f"{prefix}.self_attn",
325
+ config=config,
326
+ qk_nope_head_dim=config.qk_nope_head_dim,
327
+ qk_rope_head_dim=config.qk_rope_head_dim,
328
+ v_head_dim=config.v_head_dim,
329
+ q_lora_rank=config.q_lora_rank,
330
+ kv_lora_rank=config.kv_lora_rank,
331
+ use_nope=config.mla_use_nope,
332
+ )
333
+
334
+ if (
335
+ self.is_moe
336
+ and config.num_experts is not None
337
+ and layer_idx >= config.first_k_dense_replace
338
+ and layer_idx % config.moe_layer_freq == 0
339
+ ):
340
+ self.block_sparse_moe = KimiMoE(
341
+ config=config,
342
+ quant_config=quant_config,
343
+ prefix=f"{prefix}.mlp",
344
+ )
345
+ self.mlp = self.block_sparse_moe
346
+ else:
347
+ self.mlp = KimiMLP(
348
+ hidden_size=self.hidden_size,
349
+ intermediate_size=config.intermediate_size,
350
+ hidden_act=config.hidden_act,
351
+ quant_config=quant_config,
352
+ prefix=f"{prefix}.mlp",
353
+ )
354
+ self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
355
+ self.post_attention_layernorm = RMSNorm(
356
+ config.hidden_size, eps=config.rms_norm_eps
357
+ )
358
+
359
+ def forward(
360
+ self,
361
+ positions: torch.Tensor,
362
+ hidden_states: torch.Tensor,
363
+ residual: torch.Tensor | None,
364
+ **kwargs,
365
+ ) -> tuple[torch.Tensor, torch.Tensor]:
366
+ # Self Attention
367
+ if residual is None:
368
+ residual = hidden_states
369
+ hidden_states = self.input_layernorm(hidden_states)
370
+ else:
371
+ hidden_states, residual = self.input_layernorm(hidden_states, residual)
372
+
373
+ attn_output = torch.empty_like(hidden_states)
374
+ self.self_attn(
375
+ hidden_states=hidden_states,
376
+ positions=positions,
377
+ output=attn_output,
378
+ )
379
+ hidden_states = attn_output
380
+
381
+ # Fully Connected
382
+ hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
383
+ hidden_states = self.mlp(hidden_states)
384
+ return hidden_states, residual
385
+
386
+
387
+@support_torch_compile
388
+class KimiLinearModel(nn.Module):
389
+ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
390
+ super().__init__()
391
+
392
+ config = vllm_config.model_config.hf_text_config
393
+ model_config = vllm_config.model_config
394
+ cache_config = vllm_config.cache_config
395
+ quant_config = vllm_config.quant_config
396
+ parallel_config = vllm_config.parallel_config
397
+ self.config = config
398
+
399
+ self.padding_idx = config.pad_token_id
400
+ self.vocab_size = config.vocab_size
401
+
402
+ if get_pp_group().is_first_rank:
403
+ self.embed_tokens = VocabParallelEmbedding(
404
+ config.vocab_size,
405
+ config.hidden_size,
406
+ prefix=f"{prefix}.embed_tokens",
407
+ )
408
+ else:
409
+ self.embed_tokens = PPMissingLayer()
410
+
411
+ extra_kwargs = {}
412
+
413
+ def get_layer(prefix: str):
414
+ layer_idx = int(prefix.rsplit(".", 1)[1])
415
+ return KimiDecoderLayer(
416
+ config,
417
+ layer_idx,
418
+ cache_config,
419
+ quant_config,
420
+ parallel_config,
421
+ model_config,
422
+ prefix,
423
+ **extra_kwargs,
424
+ )
425
+
426
+ self.start_layer, self.end_layer, self.layers = make_layers(
427
+ config.num_hidden_layers,
428
+ get_layer,
429
+ prefix=f"{prefix}.layers",
430
+ )
431
+
432
+ if get_pp_group().is_last_rank:
433
+ self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
434
+ else:
435
+ self.norm = PPMissingLayer()
436
+
437
+ world_size = get_tensor_model_parallel_world_size()
438
+ assert config.num_attention_heads % world_size == 0, (
439
+ "num_attention_heads must be divisible by world_size"
440
+ )
441
+
442
+ def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
443
+ return self.embed_tokens(input_ids)
444
+
445
+ def forward(
446
+ self,
447
+ input_ids: torch.Tensor | None,
448
+ positions: torch.Tensor,
449
+ intermediate_tensors: IntermediateTensors | None,
450
+ inputs_embeds: torch.Tensor | None = None,
451
+ **kwargs,
452
+ ) -> torch.Tensor:
453
+ if get_pp_group().is_first_rank:
454
+ if inputs_embeds is not None:
455
+ hidden_states = inputs_embeds
456
+ else:
457
+ hidden_states = self.get_input_embeddings(input_ids)
458
+ residual = None
459
+ else:
460
+ assert intermediate_tensors is not None
461
+ hidden_states = intermediate_tensors["hidden_states"]
462
+ residual = intermediate_tensors["residual"]
463
+
464
+ for _, layer in enumerate(self.layers[self.start_layer : self.end_layer]):
465
+ hidden_states, residual = layer(
466
+ positions=positions,
467
+ hidden_states=hidden_states,
468
+ residual=residual,
469
+ )
470
+
471
+ if not get_pp_group().is_last_rank:
472
+ return IntermediateTensors(
473
+ {"hidden_states": hidden_states, "residual": residual}
474
+ )
475
+
476
+ hidden_states, _ = self.norm(hidden_states, residual)
477
+ return hidden_states
478
+
479
+
480
+class KimiLinearForCausalLM(
481
+ nn.Module, HasInnerState, SupportsPP, MixtureOfExperts, IsHybrid
482
+):
483
+ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
484
+ super().__init__()
485
+ self.model_config = vllm_config.model_config
486
+ self.vllm_config = vllm_config
487
+ self.config = self.model_config.hf_config
488
+ quant_config = vllm_config.quant_config
489
+ self.quant_config = quant_config
490
+ self.model = KimiLinearModel(
491
+ vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
492
+ )
493
+ if get_pp_group().is_last_rank:
494
+ self.lm_head = ParallelLMHead(
495
+ self.config.vocab_size,
496
+ self.config.hidden_size,
497
+ quant_config=quant_config,
498
+ prefix=maybe_prefix(prefix, "lm_head"),
499
+ )
500
+ else:
501
+ self.lm_head = PPMissingLayer()
502
+ logit_scale = getattr(self.config, "logit_scale", 1.0)
503
+ self.logits_processor = LogitsProcessor(
504
+ self.config.vocab_size, scale=logit_scale
505
+ )
506
+
507
+ def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
508
+ return self.model.get_input_embeddings(input_ids)
509
+
510
+ def forward(
511
+ self,
512
+ input_ids: torch.Tensor,
513
+ positions: torch.Tensor,
514
+ intermediate_tensors: IntermediateTensors | None = None,
515
+ inputs_embeds: torch.Tensor | None = None,
516
+ **kwargs,
517
+ ) -> torch.Tensor | IntermediateTensors:
518
+ hidden_states = self.model(
519
+ input_ids, positions, intermediate_tensors, inputs_embeds, **kwargs
520
+ )
521
+ return hidden_states
522
+
523
+ @classmethod
524
+ def get_mamba_state_dtype_from_config(
525
+ cls,
526
+ vllm_config: "VllmConfig",
527
+ ) -> tuple[torch.dtype, torch.dtype, torch.dtype, torch.dtype]:
528
+ return MambaStateDtypeCalculator.kda_state_dtype(
529
+ vllm_config.model_config.dtype, vllm_config.cache_config.mamba_cache_dtype
530
+ )
531
+
532
+ @classmethod
533
+ def get_mamba_state_shape_from_config(
534
+ cls, vllm_config: "VllmConfig"
535
+ ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
536
+ parallel_config = vllm_config.parallel_config
537
+ hf_config = vllm_config.model_config.hf_config
538
+ tp_size = parallel_config.tensor_parallel_size
539
+ num_spec = (
540
+ vllm_config.speculative_config.num_speculative_tokens
541
+ if vllm_config.speculative_config
542
+ else 0
543
+ )
544
+ return MambaStateShapeCalculator.kda_state_shape(
545
+ tp_size,
546
+ hf_config.linear_attn_config["num_heads"],
547
+ hf_config.linear_attn_config["head_dim"],
548
+ conv_kernel_size=hf_config.linear_attn_config["short_conv_kernel_size"],
549
+ num_spec=num_spec,
550
+ )
551
+
552
+ def compute_logits(
553
+ self,
554
+ hidden_states: torch.Tensor,
555
+ ) -> torch.Tensor | None:
556
+ return self.logits_processor(self.lm_head, hidden_states)
557
+
558
+ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
559
+ stacked_params_mapping = [
560
+ # (param_name, shard_name, shard_id)
561
+ (".gate_up_proj", ".gate_proj", 0),
562
+ (".gate_up_proj", ".up_proj", 1),
563
+ ]
564
+ if self.config.is_moe:
565
+ # Params for weights, fp8 weight scales, fp8 activation scales
566
+ # (param_name, weight_name, expert_id, shard_id)
567
+ expert_params_mapping = FusedMoE.make_expert_params_mapping(
568
+ ckpt_gate_proj_name="w1",
569
+ ckpt_down_proj_name="w2",
570
+ ckpt_up_proj_name="w3",
571
+ num_experts=self.config.num_experts,
572
+ )
573
+ else:
574
+ expert_params_mapping = []
575
+ params_dict = dict(self.named_parameters())
576
+ loaded_params: set[str] = set()
577
+ for args in weights:
578
+ name, loaded_weight = args[:2]
579
+ kwargs = args[2] if len(args) > 2 else {}
580
+ if "rotary_emb.inv_freq" in name:
581
+ continue
582
+
583
+ spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
584
+ if spec_layer is not None:
585
+ continue # skip spec decode layers for main model
586
+ if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
587
+ # Models trained using ColossalAI may include these tensors in
588
+ # the checkpoint. Skip them.
589
+ continue
590
+ for param_name, weight_name, shard_id in stacked_params_mapping:
591
+ if weight_name not in name:
592
+ continue
593
+ # We have mlp.experts[0].gate_proj in the checkpoint.
594
+ # Since we handle the experts below in expert_params_mapping,
595
+ # we need to skip here BEFORE we update the name, otherwise
596
+ # name will be updated to mlp.experts[0].gate_up_proj, which
597
+ # will then be updated below in expert_params_mapping
598
+ # for mlp.experts[0].gate_gate_up_proj, which breaks load.
599
+ if ("mlp.experts." in name) and name not in params_dict:
600
+ continue
601
+ name = name.replace(weight_name, param_name)
602
+ # Skip loading extra bias for GPTQ models.
603
+ if name.endswith(".bias") and name not in params_dict:
604
+ continue
605
+ if is_pp_missing_parameter(name, self):
606
+ continue
607
+ param = params_dict[name]
608
+ weight_loader = param.weight_loader
609
+ weight_loader(param, loaded_weight, shard_id)
610
+ break
611
+ else:
612
+ for idx, (param_name, weight_name, expert_id, shard_id) in enumerate(
613
+ expert_params_mapping
614
+ ):
615
+ if weight_name not in name:
616
+ continue
617
+ name = name.replace(weight_name, param_name)
618
+ if is_pp_missing_parameter(name, self):
619
+ continue
620
+ param = params_dict[name]
621
+ weight_loader = param.weight_loader
622
+ weight_loader(
623
+ param,
624
+ loaded_weight,
625
+ name,
626
+ expert_id=expert_id,
627
+ shard_id=shard_id,
628
+ )
629
+ break
630
+ else:
631
+ # Skip loading extra bias for GPTQ models.
632
+ if (
633
+ name.endswith(".bias")
634
+ and name not in params_dict
635
+ and not self.config.is_linear_attn
636
+ ): # noqa: E501
637
+ continue
638
+ # Remapping the name of FP8 kv-scale.
639
+ name = maybe_remap_kv_scale_name(name, params_dict)
640
+ if name is None:
641
+ continue
642
+ if is_pp_missing_parameter(name, self):
643
+ continue
644
+
645
+ param = params_dict[name]
646
+ weight_loader = getattr(
647
+ param, "weight_loader", default_weight_loader
648
+ )
649
+ weight_loader(param, loaded_weight, **kwargs)
650
+ loaded_params.add(name)
651
+
652
+
653
+def get_spec_layer_idx_from_weight_name(
654
+ config: KimiLinearConfig, weight_name: str
655
+) -> int | None:
656
+ if hasattr(config, "num_nextn_predict_layers") and (
657
+ config.num_nextn_predict_layers > 0
658
+ ):
659
+ layer_idx = config.num_hidden_layers
660
+ for i in range(config.num_nextn_predict_layers):
661
+ if weight_name.startswith(f"model.layers.{layer_idx + i}."):
662
+ return layer_idx + i
663
+ return None
vllm/model_executor/models/registry.py
+1
@@ -118,6 +118,7 @@ _TEXT_GENERATION_MODELS = {
118
"InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
119
"JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
120
"JambaForCausalLM": ("jamba", "JambaForCausalLM"),
121
+ "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"), # noqa: E501
122
"Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
123
"Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
124
"LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
vllm/transformers_utils/config.py
+1
@@ -79,6 +79,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
79
deepseek_v3="DeepseekV3Config",
80
deepseek_v32="DeepseekV3Config",
81
flex_olmo="FlexOlmoConfig",
82
+ kimi_linear="KimiLinearConfig",
83
kimi_vl="KimiVLConfig",
84
Llama_Nemotron_Nano_VL="Nemotron_Nano_VL_Config",
85
RefinedWeb="RWConfig", # For tiiuae/falcon-40b(-instruct)
vllm/transformers_utils/configs/__init__.py
+2
@@ -19,6 +19,7 @@ from vllm.transformers_utils.configs.eagle import EAGLEConfig
19
from vllm.transformers_utils.configs.falcon import RWConfig
20
from vllm.transformers_utils.configs.flex_olmo import FlexOlmoConfig
21
from vllm.transformers_utils.configs.jais import JAISConfig
22
+from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig
23
from vllm.transformers_utils.configs.kimi_vl import KimiVLConfig
24
from vllm.transformers_utils.configs.lfm2_moe import Lfm2MoeConfig
25
from vllm.transformers_utils.configs.medusa import MedusaConfig
@@ -54,6 +55,7 @@ __all__ = [
55
"MiDashengLMConfig",
56
"MLPSpeculatorConfig",
57
"MoonViTConfig",
58
+ "KimiLinearConfig",
59
"KimiVLConfig",
60
"NemotronConfig",
61
"NemotronHConfig",
vllm/transformers_utils/configs/kimi_linear.py
new
+144
@@ -0,0 +1,144 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+
4
+from transformers.configuration_utils import PretrainedConfig
5
+
6
+from vllm.logger import init_logger
7
+
8
+logger = init_logger(__name__)
9
+
10
+
11
+class KimiLinearConfig(PretrainedConfig):
12
+ model_type = "kimi_linear"
13
+ keys_to_ignore_at_inference = ["past_key_values"]
14
+
15
+ def __init__(
16
+ self,
17
+ model_type="kimi_linear",
18
+ vocab_size=163840,
19
+ hidden_size=4096,
20
+ head_dim=None,
21
+ intermediate_size=11008,
22
+ num_hidden_layers=32,
23
+ num_attention_heads=32,
24
+ num_key_value_heads=None,
25
+ hidden_act="silu",
26
+ initializer_range=0.02,
27
+ rms_norm_eps=1e-6,
28
+ use_cache=True,
29
+ pad_token_id=0,
30
+ bos_token_id=1,
31
+ eos_token_id=2,
32
+ rope_theta=10000.0,
33
+ rope_scaling=None,
34
+ tie_word_embeddings=False,
35
+ moe_intermediate_size: int | None = None,
36
+ moe_renormalize: bool = True,
37
+ moe_router_activation_func: str = "sigmoid",
38
+ num_experts: int | None = None,
39
+ num_experts_per_token: int | None = None,
40
+ num_shared_experts: int = 0,
41
+ routed_scaling_factor: float = 1.0,
42
+ first_k_dense_replace: int = 0,
43
+ moe_layer_freq: int = 1,
44
+ use_grouped_topk: bool = True,
45
+ num_expert_group: int = 1,
46
+ topk_group: int = 1,
47
+ q_lora_rank: int | None = None,
48
+ kv_lora_rank: int | None = None,
49
+ qk_nope_head_dim: int | None = None,
50
+ qk_rope_head_dim: int | None = None,
51
+ v_head_dim: int | None = None,
52
+ mla_use_nope: bool | None = False,
53
+ num_nextn_predict_layers: int = 0,
54
+ linear_attn_config: dict | None = None,
55
+ **kwargs,
56
+ ):
57
+ self.model_type = model_type
58
+ self.vocab_size = vocab_size
59
+ self.hidden_size = hidden_size
60
+ self.head_dim = (
61
+ head_dim if head_dim is not None else hidden_size // num_attention_heads
62
+ )
63
+ self.intermediate_size = intermediate_size
64
+ self.num_hidden_layers = num_hidden_layers
65
+ self.num_attention_heads = num_attention_heads
66
+
67
+ # for backward compatibility
68
+ if num_key_value_heads is None:
69
+ num_key_value_heads = num_attention_heads
70
+
71
+ self.num_key_value_heads = num_key_value_heads
72
+ self.hidden_act = hidden_act
73
+ self.initializer_range = initializer_range
74
+ self.rms_norm_eps = rms_norm_eps
75
+ self.use_cache = use_cache
76
+ self.rope_theta = rope_theta
77
+ self.rope_scaling = rope_scaling
78
+
79
+ self.q_lora_rank = q_lora_rank
80
+ self.kv_lora_rank = kv_lora_rank
81
+ self.qk_nope_head_dim = qk_nope_head_dim
82
+ self.qk_rope_head_dim = qk_rope_head_dim
83
+ self.v_head_dim = v_head_dim
84
+ self.mla_use_nope = mla_use_nope
85
+ # moe config
86
+ self.num_experts = num_experts
87
+ self.num_experts_per_token = num_experts_per_token
88
+ self.moe_renormalize = moe_renormalize
89
+ self.num_shared_experts = num_shared_experts
90
+ self.routed_scaling_factor = routed_scaling_factor
91
+ self.moe_router_activation_func = moe_router_activation_func
92
+ assert self.moe_router_activation_func in ("softmax", "sigmoid")
93
+ self.moe_intermediate_size = moe_intermediate_size
94
+ self.first_k_dense_replace = first_k_dense_replace
95
+ self.moe_layer_freq = moe_layer_freq
96
+ self.use_grouped_topk = use_grouped_topk
97
+ self.num_expert_group = num_expert_group
98
+ self.topk_group = topk_group
99
+ self.num_nextn_predict_layers = num_nextn_predict_layers
100
+
101
+ if linear_attn_config is not None:
102
+ assert linear_attn_config["kda_layers"] is not None
103
+ assert linear_attn_config["full_attn_layers"] is not None
104
+ self.linear_attn_config = linear_attn_config
105
+
106
+ super().__init__(
107
+ pad_token_id=pad_token_id,
108
+ bos_token_id=bos_token_id,
109
+ eos_token_id=eos_token_id,
110
+ tie_word_embeddings=tie_word_embeddings,
111
+ **kwargs,
112
+ )
113
+
114
+ @property
115
+ def is_mla(self):
116
+ return (
117
+ self.q_lora_rank is not None
118
+ or self.kv_lora_rank is not None
119
+ or self.qk_nope_head_dim is not None
120
+ or self.qk_rope_head_dim is not None
121
+ or self.v_head_dim is not None
122
+ or self.mla_use_nope is True
123
+ )
124
+
125
+ @property
126
+ def is_moe(self):
127
+ return self.num_experts is not None
128
+
129
+ @property
130
+ def is_linear_attn(self) -> bool:
131
+ return not (
132
+ self.linear_attn_config is None
133
+ or (
134
+ isinstance(self.linear_attn_config, dict)
135
+ and self.linear_attn_config["kda_layers"] is not None
136
+ and len(self.linear_attn_config["kda_layers"]) == 0
137
+ )
138
+ )
139
+
140
+ def is_kda_layer(self, layer_idx: int):
141
+ return (
142
+ self.linear_attn_config is not None
143
+ and (layer_idx + 1) in self.linear_attn_config["kda_layers"]
144
+ )
vllm/v1/worker/gpu_model_runner.py
+12
-19
@@ -8,6 +8,7 @@ from collections import defaultdict
8
from collections.abc import Iterator
9
from contextlib import contextmanager
10
from copy import deepcopy
11
+from functools import reduce
12
from itertools import product
13
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
14
@@ -4134,26 +4135,18 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
4135
4136
def calculate_reorder_batch_threshold(self) -> None:
4137
"""
4137
- Check that if any backends reorder batches; that the reordering
4138
- is compatible (e.g., decode threshold is the same)
4138
+ Choose the minimum reorder batch threshold from all attention groups.
4139
+ Backends should be able to support lower threshold then what they request
4140
+ just may have a performance penalty due to that backend treating decodes
4141
+ as prefills.
4142
"""
4140
- for group in self._attn_group_iterator():
4141
- attn_metadata_builder_i = group.get_metadata_builder()
4142
-
4143
- # check that if any backends reorder batches; that the reordering
4144
- # is compatible (e.g., decode threshold is the same)
4145
- reorder_batch_threshold_i = attn_metadata_builder_i.reorder_batch_threshold
4146
- if reorder_batch_threshold_i is not None:
4147
- if self.reorder_batch_threshold is not None:
4148
- if reorder_batch_threshold_i != self.reorder_batch_threshold:
4149
- raise ValueError(
4150
- f"Attention backend reorders decodes with "
4151
- f"threshold {reorder_batch_threshold_i} but other "
4152
- f"backend uses threshold "
4153
- f"{self.reorder_batch_threshold}"
4154
- )
4155
- else:
4156
- self.reorder_batch_threshold = reorder_batch_threshold_i
4143
+ min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)
4144
+
4145
+ reorder_batch_thresholds = [
4146
+ group.get_metadata_builder().reorder_batch_threshold
4147
+ for group in self._attn_group_iterator()
4148
+ ]
4149
+ self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds)
4150
4151
def _find_compatible_block_sizes(
4152
self,