EngineX-Cambricon/enginex-mlu370-vllm
10
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

add deepseekv3 and llama4

Browse Source
This commit is contained in:
Chranos
2026-02-11 16:03:06 +08:00
parent 1f77771852
commit 5457f79dbb
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm-v0.6.2/vllm_mlu/vllm_mlu/config.py
View File

@@ -37,7 +37,7 @@ def vllm__config__CacheConfig___verify_cache_dtype(self) -> None:
def vllm__config__ModelConfig__get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int: def vllm__config__ModelConfig__get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
"""Returns the number of KV heads per GPU.""" """Returns the number of KV heads per GPU."""
if hasattr(self.hf_text_config,"model_type") and self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3'): if hasattr(self.hf_text_config,"model_type") and self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp'):
# feature flag MLA # feature flag MLA
return 1 return 1
total_num_kv_heads = self.get_total_num_kv_heads() total_num_kv_heads = self.get_total_num_kv_heads()
@@ -51,7 +51,7 @@ def vllm__config__ModelConfig__get_num_kv_heads(self, parallel_config: "Parallel
def vllm__config__ModelConfig__get_head_size(self) -> int: def vllm__config__ModelConfig__get_head_size(self) -> int:
# TODO remove hard code # TODO remove hard code
if hasattr(self.hf_text_config, "model_type" if hasattr(self.hf_text_config, "model_type"
) and self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3'): ) and self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp'):
''' '''
============================= =============================
Modify by vllm_mlu Modify by vllm_mlu
@@ -109,7 +109,7 @@ def vllm__config__LoRAConfig__verify_with_model_config(self, model_config: Model
def vllm__config__ModelConfig__is_deepseek_v2(self) -> bool: def vllm__config__ModelConfig__is_deepseek_v2(self) -> bool:
result = hasattr( result = hasattr(
self.hf_text_config, self.hf_text_config,
"model_type") and self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3') "model_type") and self.hf_text_config.model_type in ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp')
return result return result
MluHijackObject.apply_hijack(ModelConfig, MluHijackObject.apply_hijack(ModelConfig,