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

[Misc] Upgrade vllm hash to 12_14 (#5000)

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

### Does this PR introduce _any_ user-facing change?
1. fix https://github.com/vllm-project/vllm/pull/27938
2. fix https://github.com/vllm-project/vllm/pull/27145
pooling models now supports chunked prefill and prefix caching,
3. fix https://github.com/vllm-project/vllm/pull/30181
define the CPU fields in the field config where they really belong.
4. fix https://github.com/vllm-project/vllm/pull/28168
define the CPU fields in the field config where they really belong.
5. fix https://github.com/vllm-project/vllm/pull/30201
some moudle rename
6. fix https://github.com/vllm-project/vllm/pull/29067
fusedmoe moudle refactor
7. fix https://github.com/vllm-project/vllm/pull/29066
fusedmoe moudle refactor
8. fix https://github.com/vllm-project/vllm/pull/29624
### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: wangli <wangli858794774@gmail.com>
This commit is contained in:
Li Wang
2025-12-15 19:54:23 +08:00
committed by GitHub
parent 3b7eb5179f
commit 8d2998d0e4
17 changed files with 167 additions and 1183 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
vllm_ascend/ops/fused_moe/fused_moe.py
View File

@@ -153,7 +153,7 @@ class AscendFusedMoE(FusedMoE):
AscendFusedMoE.moe_counter += 1
self.moe_instance_id = AscendFusedMoE.moe_counter
self.expert_map = None
self._expert_map = None
self.log2phy = None
if self.quant_config is None:
@@ -184,7 +184,7 @@ class AscendFusedMoE(FusedMoE):
dtype=vllm_config.model_config.dtype)
# init moe.
self.local_num_experts, self.expert_map, _ = determine_expert_map(
self.local_num_experts, self._expert_map, _ = determine_expert_map(
self.ep_size, self.ep_rank, self.global_num_experts)
# TODO: Temporary flag to indicate if static EPLB is enabled. This is a
# workaround to bypass a quantization check that fails with float weights.
@@ -200,7 +200,7 @@ class AscendFusedMoE(FusedMoE):
self.expert_load_balancer.get_global_redundant_expert_num())
self.global_num_experts = num_experts + self.global_redundant_expert_num
try:
self.local_num_experts, self.expert_map = (
self.local_num_experts, self._expert_map = (
self.expert_load_balancer.get_rank_placement_map(
self.moe_instance_id, self.ep_rank))
self.log2phy = self.expert_load_balancer.get_rank_log2phy_map(
@@ -216,16 +216,16 @@ class AscendFusedMoE(FusedMoE):
if self.dynamic_eplb:
self.log2phy = determine_default_log2phy_map(
self.global_num_experts, self.ep_size, self.ep_rank).npu()
if self.expert_map is not None and isinstance(self.expert_map,
torch.Tensor):
if self._expert_map is not None and isinstance(self._expert_map,
torch.Tensor):
logger.info_once(
"[EP Rank %s/%s] Expert parallelism is enabled. Local/global"
" number of experts: %s/%s. Experts local to global index map:"
" %s.", self.ep_rank, self.ep_size, self.local_num_experts,
self.global_num_experts,
get_compressed_expert_map(self.expert_map))
get_compressed_expert_map(self._expert_map))
local_num_experts = (torch.sum(
self.expert_map != -1) if self.expert_map is not None else
self._expert_map != -1) if self._expert_map is not None else
self.global_num_experts)
if self.dynamic_eplb:
self.moe_load = torch.zeros(local_num_experts,
@@ -276,10 +276,16 @@ class AscendFusedMoE(FusedMoE):
return QuantType.NONE
def update_expert_map(self, new_expert_map):
self.expert_map = new_expert_map
self._expert_map = new_expert_map
def get_map(self):
return self.expert_map
@property
def expert_map(self) -> torch.Tensor | None:
return self._expert_map
@expert_map.setter
def expert_map(self, new_expert_map):
# TODO(Potabk): Remove this once we drop vllm v0.12.0(This makes backward compatibility with vllm v0.12.0)
self._expert_map = new_expert_map
def get_log2phy_map(self):
return self.log2phy