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

[Bugfix] Fix the graph capture failure issue in the eagle3+full scenario. (#5553)

Browse Source 
### What this PR does / why we need it?
When launching the service in the scenario where the
cudagraph_mode is set to FULL and Eagle3 acceleration is enabled for
inference, an error in fia will cause graph capture to fail. This PR
fixes the issue.

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

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
7157596103

Signed-off-by: WithHades <244036962@qq.com>
This commit is contained in:
无脸男
2026-01-07 15:57:16 +08:00
committed by GitHub
parent 2b8a9ce8bd
commit 1140789e83
2 changed files with 10 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm_ascend/spec_decode/eagle_proposer.py
View File

@@ -73,7 +73,8 @@ class EagleProposer(VllmEagleProposer):
self.pcp_size = self.runner.pcp_size self.pcp_size = self.runner.pcp_size
self.decode_threshold = 1 + self.num_speculative_tokens self.decode_threshold = 1 + self.num_speculative_tokens
self.query_start_loc = self.runner._make_buffer(
self.runner.max_num_reqs + 1, dtype=torch.int32)
self.arange_cpu = torch.arange(self.arange.shape[0], self.arange_cpu = torch.arange(self.arange.shape[0],
device="cpu", device="cpu",
dtype=torch.int32) dtype=torch.int32)
@@ -200,10 +201,14 @@ class EagleProposer(VllmEagleProposer):
num_computed_tokens_cpu = ( num_computed_tokens_cpu = (
self.runner.input_batch. self.runner.input_batch.
num_computed_tokens_cpu_tensor[:num_reqs]) num_computed_tokens_cpu_tensor[:num_reqs])
self.query_start_loc.cpu[:num_reqs + 1] = torch.tensor(
[0] + self.runner.actual_seq_lengths_q[:num_reqs],
device="cpu",
dtype=torch.int32)
self.query_start_loc.copy_to_gpu()
common_attn_metadata = AscendCommonAttentionMetadata( common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=self.runner.query_start_loc.gpu[:num_reqs + 1], query_start_loc=self.query_start_loc.gpu[:num_reqs + 1],
query_start_loc_cpu=self.runner.query_start_loc.cpu[:num_reqs + query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs + 1],
1],
seq_lens_cpu=self.runner.seq_lens.cpu, seq_lens_cpu=self.runner.seq_lens.cpu,
seq_lens=self.runner.seq_lens.gpu[:num_reqs], seq_lens=self.runner.seq_lens.gpu[:num_reqs],
num_reqs=num_reqs, num_reqs=num_reqs,

9
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1939,14 +1939,7 @@ class NPUModelRunner(GPUModelRunner):
[0] * dcp_world_size for _ in range(pcp_world_size) [0] * dcp_world_size for _ in range(pcp_world_size)
] for _ in range(num_tokens)] ] for _ in range(num_tokens)]
long_seq_metadata.num_computed_tokens_of_pcp_dcp = num_computed_tokens_of_pcp_dcp long_seq_metadata.num_computed_tokens_of_pcp_dcp = num_computed_tokens_of_pcp_dcp
# QUESTION: Why do we separately set query_start_loc for spec in the first place?
# While in _prepare_inputs we don't?
if self.speculative_config:
self.query_start_loc.cpu[:num_reqs + 1] = torch.tensor(
[0] + self.actual_seq_lengths_q[:num_reqs],
device="cpu",
dtype=torch.int32)
self.query_start_loc.copy_to_gpu()
common_attn_metadata = AscendCommonAttentionMetadata( common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=self.query_start_loc.gpu[:num_reqs + 1], query_start_loc=self.query_start_loc.gpu[:num_reqs + 1],
query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs + query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs +