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

[Bugfix] fix hash conflict due to reset incompatible configuations (#6368)

Browse Source 
### What this PR does / why we need it?
[Bugfix] fix hash conflict due to reset incompatible configuations

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

### How was this patch tested?

- vLLM version: v0.14.1
- vLLM main:
dc917cceb8

Signed-off-by: hfadzxy <starmoon_zhang@163.com>
This commit is contained in:
zhangxinyuehfad
2026-02-03 10:32:02 +08:00
committed by GitHub
parent b6256e8bc9
commit 05cc03d785
1 changed files with 8 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
vllm_ascend/platform.py
View File

@@ -658,20 +658,8 @@ class NPUPlatform(Platform):
)
model_config.disable_cascade_attn = False
# ==================== 2. Parallel Config ====================
if vllm_config.parallel_config:
# Only allow the default all2all backend; others like deepep are not supported
default_backend = "allgather_reducescatter"
current_backend = getattr(vllm_config.parallel_config, "all2all_backend", default_backend)
if current_backend != default_backend:
logger.warning(
"Parameter '--all2all-backend' is set to '%s', which may be "
"incompatible with Ascend. Using internal plugin mechanisms.",
current_backend,
)
vllm_config.parallel_config.all2all_backend = default_backend
# ==================== 3. Cache Config ====================
# ==================== 2. Cache Config ====================
if vllm_config.cache_config:
# Check and reset cpu_kvcache_space_bytes
if getattr(vllm_config.cache_config, "cpu_kvcache_space_bytes", False):
logger.warning(
@@ -679,7 +667,7 @@ class NPUPlatform(Platform):
)
vllm_config.cache_config.cpu_kvcache_space_bytes = None
# ==================== 4. MultiModal Config ====================
# ==================== 3. MultiModal Config ====================
multimodal_config = getattr(model_config, "multimodal_config", None) if model_config else None
if multimodal_config:
# Ascend uses a different mechanism for Multi-Modal attention
@@ -690,7 +678,7 @@ class NPUPlatform(Platform):
)
multimodal_config.mm_encoder_attn_backend = None
# ==================== 5. Observability Config ====================
# ==================== 4. Observability Config ====================
if vllm_config.observability_config:
# NVTX tracing is NVIDIA specific
if getattr(vllm_config.observability_config, "enable_layerwise_nvtx_tracing", False):
@@ -700,7 +688,7 @@ class NPUPlatform(Platform):
)
vllm_config.observability_config.enable_layerwise_nvtx_tracing = False
# ==================== 6. Scheduler Config ====================
# ==================== 5. Scheduler Config ====================
if vllm_config.scheduler_config:
# Partial prefills are specific to ROCm optimization
if getattr(vllm_config.scheduler_config, "max_num_partial_prefills", 1) != 1:
@@ -709,7 +697,7 @@ class NPUPlatform(Platform):
)
vllm_config.scheduler_config.max_num_partial_prefills = 1
# ==================== 7. Speculative Config ====================
# ==================== 6. Speculative Config ====================
if vllm_config.speculative_config:
# Ascend automatically inherits main model quantization
if getattr(vllm_config.speculative_config, "quantization", None) is not None:
@@ -719,7 +707,7 @@ class NPUPlatform(Platform):
)
vllm_config.speculative_config.quantization = None
# ==================== 8. KV Transfer Config ====================
# ==================== 7. KV Transfer Config ====================
if vllm_config.kv_transfer_config:
# Buffer size is primarily tied to NCCL (GPU) backends
current_buffer_size = getattr(vllm_config.kv_transfer_config, "kv_buffer_size", 1e9)
@@ -739,7 +727,7 @@ class NPUPlatform(Platform):
)
vllm_config.kv_transfer_config.enable_permute_local_kv = False
# ==================== 9. Attention Config ====================
# ==================== 8. Attention Config ====================
if vllm_config.attention_config:
att_config = vllm_config.attention_config