ZIBEE/bi_150-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Upgrade to vllm 0.17.0 corex v4.1 overlay

Browse Source
This commit is contained in:
LZiBee
2026-04-29 19:38:22 +08:00
parent 8fac6062e4
commit 938d0854a5
430 changed files with 35969 additions and 14511 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
vllm/model_executor/models/qwen3_5.py
View File

@@ -274,7 +274,6 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
1,
1,
config.hidden_size,
dtype=config.dtype,
),
)
self.ffn_layer_scale = torch.nn.Parameter(
@@ -282,7 +281,6 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
1,
1,
config.hidden_size,
dtype=config.dtype,
),
)
@@ -444,11 +442,7 @@ class Qwen3_5Model(Qwen3NextModel):
# qwen3.5 no need to transpose
# loaded_weight = loaded_weight.transpose(-1, -2)
if "experts.gate_up_proj" in name:
if loaded_weight.shape[-2] != 1:
chunk_dim = -2
else:
chunk_dim = -1
loaded_weight = loaded_weight.chunk(2, dim=chunk_dim)
loaded_weight = loaded_weight.chunk(2, dim=-2)
success_w1 = self.load_fused_expert_weights(
name_mapped,
params_dict,
@@ -544,6 +538,7 @@ class Qwen3_5ForCausalLMBase(
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
scheduler_config = vllm_config.scheduler_config
if cache_config.mamba_cache_mode == "all":
raise NotImplementedError(
@@ -633,6 +628,9 @@ class Qwen3_5MoeForCausalLM(Qwen3_5ForCausalLMBase, QwenNextMixtureOfExperts):
dummy_inputs=Qwen3VLDummyInputsBuilder,
)
class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid):
# Qwen3.5 does not support multimodal pruning (EVS).
supports_multimodal_pruning = False
packed_modules_mapping = Qwen3VLForConditionalGeneration.packed_modules_mapping | {
"in_proj_qkvz": ["in_proj_qkv", "in_proj_z"],
"in_proj_ba": ["in_proj_b", "in_proj_a"],
@@ -648,10 +646,8 @@ class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid)
self.config = config
self.multimodal_config = multimodal_config
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
self.video_pruning_rate = multimodal_config.video_pruning_rate
self.is_multimodal_pruning_enabled = (
multimodal_config.is_multimodal_pruning_enabled()
)
# Qwen3.5 does not support multimodal pruning (EVS).
self.is_multimodal_pruning_enabled = False
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.visual = Qwen3_VisionTransformer(
@@ -698,6 +694,12 @@ class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid)
return inputs_embeds
def recompute_mrope_positions(self, *args, **kwargs):
raise NotImplementedError(
"Qwen3.5 does not support multimodal pruning (EVS). "
"recompute_mrope_positions should never be called."
)
def forward(
self,
input_ids: torch.Tensor,
@@ -856,10 +858,8 @@ class Qwen3_5MoeForConditionalGeneration(
self.config = config
self.multimodal_config = multimodal_config
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
self.video_pruning_rate = multimodal_config.video_pruning_rate
self.is_multimodal_pruning_enabled = (
multimodal_config.is_multimodal_pruning_enabled()
)
# Qwen3.5 does not support multimodal pruning (EVS).
self.is_multimodal_pruning_enabled = False
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.visual = Qwen3_VisionTransformer(