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

[Bugfix] Obtain kernel block size for computing slot mapping correctly (#7019)

Browse Source 
### What this PR does / why we need it?
This PR aims to fix incorrect slot mapping in qwen35 due to mismatched
block size. In qwen35, we should use `kernel_block_size` so that we can
compute it in a correct way, and it is obtained in `load_model` when we
have a chance to grab `draft_attn_layers`.

- vLLM version: v0.16.0
- vLLM main:
15d76f74e2

Signed-off-by: Zetong Li <slippersss@126.com>
This commit is contained in:
Zetong Li
2026-03-09 11:05:01 +08:00
committed by GitHub
parent a3f4f6b10b
commit 06ec136f08
1 changed files with 12 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm_ascend/spec_decode/eagle_proposer.py
View File

@@ -161,13 +161,19 @@ class AscendEagleProposer(EagleProposer):
) )
indexer_layers = get_layers_from_vllm_config(self.vllm_config, DeepseekV32IndexerCache).keys() indexer_layers = get_layers_from_vllm_config(self.vllm_config, DeepseekV32IndexerCache).keys()
draft_attn_layer = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase).keys() draft_attn_layers_dict = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
draft_attn_layers = draft_attn_layers_dict.keys()
draft_attn_layer_names = draft_attn_layer - target_attn_layer_names draft_attn_layer_names = draft_attn_layers - target_attn_layer_names
draft_indexer_layer_names = indexer_layers - target_indexer_layer_names draft_indexer_layer_names = indexer_layers - target_indexer_layer_names
draft_attn_layer_names = draft_attn_layer_names - draft_indexer_layer_names draft_attn_layer_names = draft_attn_layer_names - draft_indexer_layer_names
assert len(draft_attn_layer_names) == 1 assert len(draft_attn_layer_names) == 1
self.attn_layer_names = list(sorted(draft_attn_layer_names)) self.attn_layer_names = list(sorted(draft_attn_layer_names))
self.kernel_block_size = (
draft_attn_layers_dict[self.attn_layer_names[0]].get_attn_backend().get_supported_kernel_block_sizes()[0]
)
self.piece_all_attn_layer_name = [] self.piece_all_attn_layer_name = []
for _ in range(self.num_speculative_tokens): for _ in range(self.num_speculative_tokens):
self.piece_all_attn_layer_name.append([name for name in self.attn_layer_names]) self.piece_all_attn_layer_name.append([name for name in self.attn_layer_names])
@@ -978,7 +984,10 @@ class AscendEagleProposer(EagleProposer):
slot_mapping = mtp_slot_mapping[slot_indices] slot_mapping = mtp_slot_mapping[slot_indices]
common_attn_metadata.slot_mapping[: batch_size * self.pcp_size] = slot_mapping common_attn_metadata.slot_mapping[: batch_size * self.pcp_size] = slot_mapping
else: else:
block_size = attn_metadata_builder.kv_cache_spec.block_size # NOTE: In vllm, `block_size = attn_metadata_builder.kv_cache_spec.block_size`.
# However, in vllm-ascend, the above value can be multiple of `kernel_block_size`,
# which is not correct for computing `slot_mapping` below.
block_size = self.kernel_block_size
# Compute the slot mapping. # Compute the slot mapping.
if self.uses_mrope: if self.uses_mrope: