EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[MM][Bugfix] Update hf_config to hf_text_config (#5319)

Browse Source 
### What this PR does / why we need it?

Following https://github.com/vllm-project/vllm-ascend/pull/5205, update
`hf_config` to `hf_text_config`.

Find more details at
https://github.com/vllm-project/vllm-ascend/pull/5205#issuecomment-3675417534
and
https://github.com/vllm-project/vllm-ascend/pull/5205#issuecomment-3677920872.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef

Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen
2026-01-06 16:41:39 +08:00
committed by GitHub
parent 293b2275df
commit b94d589769
23 changed files with 44 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_ascend/utils.py
View File

@@ -468,7 +468,7 @@ def update_default_aclgraph_sizes(vllm_config: VllmConfig) -> None:
# on special shapes.
# TODO(Angazenn): we will remove this once _npu_paged_attention is fully
# replaced by npu_fused_infer_attention_score which does not contain such bugs.
if vllm_config.model_config and vllm_config.model_config.hf_config.model_type == "qwen3_moe" \
if vllm_config.model_config and vllm_config.model_config.hf_text_config.model_type == "qwen3_moe" \
and vllm_config.parallel_config.tensor_parallel_size == 1 \
and vllm_config.parallel_config.data_parallel_size > 1 :
@@ -503,7 +503,7 @@ def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
)
return
hf_config = vllm_config.model_config.hf_config
hf_config = vllm_config.model_config.hf_text_config
if hasattr(hf_config, 'num_hidden_layers'):
num_hidden_layers = hf_config.num_hidden_layers
else:
@@ -826,7 +826,7 @@ def is_moe_model(vllm_config: VllmConfig):
"""Checks if the model is a MoE model by config"""
global _IS_MOE_MODEL
if _IS_MOE_MODEL is None:
model_configs = vllm_config.model_config.hf_config.to_dict()
model_configs = vllm_config.model_config.hf_text_config.to_dict()
_IS_MOE_MODEL = _is_contain_expert(model_configs)
return _IS_MOE_MODEL
@@ -842,7 +842,7 @@ def speculative_enable_dispatch_gmm_combine_decode(
if speculative_method in ["eagle", "eagle3"]:
return False
if speculative_method == "mtp":
mtp_quant_type = getattr(vllm_config.model_config.hf_config,
mtp_quant_type = getattr(vllm_config.model_config.hf_text_config,
"mtp_quantize", None)
return mtp_quant_type == "w8a8_dynamic"
return False
@@ -875,7 +875,7 @@ def has_rope(vllm_config: VllmConfig):
"""Checks if the model uses rope."""
global _HAS_ROPE
if _HAS_ROPE is None and vllm_config and vllm_config.model_config:
hf_config = vllm_config.model_config.hf_config.to_dict()
hf_config = vllm_config.model_config.hf_text_config.to_dict()
_HAS_ROPE = "rope_parameters" in hf_config
return _HAS_ROPE
@@ -1091,7 +1091,7 @@ def refresh_block_size(vllm_config):
return
# TODO(MengqingCao): Remove the model_type check, after resolving the hidden error in get_kv_cache_groups.
if not model_config.hf_config.model_type == "qwen3_next" and cache_config.block_size != 128:
if not model_config.hf_text_config.model_type == "qwen3_next" and cache_config.block_size != 128:
if cache_config.enable_prefix_caching or scheduler_config.enable_chunked_prefill:
logger.info(
"Block size is set to 128 if prefix cache or chunked prefill is enabled."