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

[Bugfix] Fix zero attention output in qwen3-next (#3572)

Browse Source 
### What this PR does / why we need it?
Since Attention and LinearAttention share the same ```slot_mapping```,
and the ```slot_mapping``` for LinearAttention is all zeros, the
```slot_mapping``` for Attention gets overwritten, resulting in the
computed output being all zeros.

This PR removes the uniformly managed ```self.slot_mapping``` and
directly passes the ```slot_mapping``` from ```input_batch.blocktable```
to ```attn_metadata```, along with modifying the relevant references.
Due to hardware, the data type of ```block_table.slot_mapping``` needs
to be set to int32.

### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed with existing test.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: QilaiZhang <245706640@qq.com>
This commit is contained in:
QilaiZhang
2025-10-25 09:47:03 +08:00
committed by GitHub
parent e33751ef8b
commit d30bb95b90
3 changed files with 13 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm_ascend/spec_decode/eagle_proposer.py
View File

@@ -359,7 +359,8 @@ class EagleProposer(Proposer):
actual_seq_lengths_q=self.runner.actual_seq_lengths_q,
block_table_tensor=self.runner.input_batch.block_table[0].
get_device_tensor(),
slot_mapping=self.runner.slot_mapping,
slot_mapping=self.runner.input_batch.block_table[0].
slot_mapping,
positions=self.runner.positions,
attn_mask=self.runner.attn_mask,
spec_attn_mask=self.runner.spec_attn_mask,

2
vllm_ascend/worker/block_table.py
View File

@@ -83,7 +83,7 @@ class BlockTable:
pin_memory=self.pin_memory)
self.slot_mapping_np = self.slot_mapping_cpu.numpy()
self.slot_mapping = torch.zeros(self.max_num_batched_tokens,
dtype=torch.int64,
dtype=torch.int32,
device=self.device)
try:
self.pcp_world_size = get_pcp_group(

29
vllm_ascend/worker/model_runner_v1.py
View File

@@ -404,9 +404,6 @@ class NPUModelRunner(LoRAModelRunnerMixin):
self.seq_lens = torch.zeros(self.max_num_reqs,
dtype=torch.int32,
device=self.device)
self.slot_mapping = torch.zeros(self.max_num_tokens,
dtype=torch.int32,
device=self.device)
if self.vllm_config.model_config.use_mla and \
self.compilation_config.cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY:
@@ -470,11 +467,6 @@ class NPUModelRunner(LoRAModelRunnerMixin):
pin_memory=True)
self.positions_np = self.positions_cpu.numpy()
self.slot_mapping_cpu = torch.zeros(self.max_num_tokens,
dtype=torch.int32,
device="cpu",
pin_memory=True)
self.slot_mapping_np = self.slot_mapping_cpu.numpy()
self.query_start_loc_cpu = torch.zeros(self.max_num_reqs + 1,
dtype=torch.int32,
device="cpu",
@@ -1673,12 +1665,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
else:
blk_table = self.input_batch.block_table[kv_cache_group_id]
blk_table_tensor = blk_table.get_device_tensor()
slot_mapping = blk_table.slot_mapping_cpu[:slot_mapping_size]
self.slot_mapping[:slot_mapping_size].copy_(
slot_mapping[:slot_mapping_size],
non_blocking=True,
)
self.slot_mapping[slot_mapping_size:].fill_(0)
slot_mapping = blk_table.slot_mapping[:slot_mapping_size]
blk_table.slot_mapping[slot_mapping_size:].fill_(0)
if self.pcp_size > 1:
assert pcp_unpad_mask is not None
pcp_padded_slot_mapping = self.pcp_padded_slot_mapping[:
@@ -1688,9 +1676,10 @@ class NPUModelRunner(LoRAModelRunnerMixin):
0]]
pcp_padded_slot_mapping.fill_(-1)
pcp_padded_slot_mapping[
pcp_unpad_mask] = self.slot_mapping[:slot_mapping_size]
self.slot_mapping[:long_seq_metadata.
num_actual_tokens_pcp_padded] = pcp_padded_slot_mapping
pcp_unpad_mask] = blk_table.slot_mapping[:
slot_mapping_size]
blk_table.slot_mapping[:long_seq_metadata.
num_actual_tokens_pcp_padded] = pcp_padded_slot_mapping
# Make AscendCommonAttentionMetadata
common_attn_metadata = AscendCommonAttentionMetadata(
@@ -1704,7 +1693,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
actual_seq_lengths_q=self.actual_seq_lengths_q,
# TODO: change this to the right block table for linear attn
block_table_tensor=blk_table_tensor[:num_reqs],
slot_mapping=self.slot_mapping,
slot_mapping=slot_mapping,
num_computed_tokens_cpu=num_computed_tokens_cpu,
positions=self.positions,
attn_mask=self.attn_mask,
@@ -2517,6 +2506,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
self.kv_cache_config.kv_cache_groups):
block_table_tensor = self.input_batch.block_table[
kv_cache_group_id].get_device_tensor()
slot_mapping = self.input_batch.block_table[
kv_cache_group_id].slot_mapping
common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=torch.tensor(
[0] + self.actual_seq_lengths_q[:num_reqs],
@@ -2530,7 +2521,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
num_actual_tokens=num_tokens,
actual_seq_lengths_q=self.actual_seq_lengths_q,
block_table_tensor=block_table_tensor[:num_reqs],
slot_mapping=self.slot_mapping,
slot_mapping=slot_mapping,
num_computed_tokens_cpu=num_computed_tokens_cpu,
positions=self.positions,
attn_mask=self.attn_mask,