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

[BugFix] Adapted Qwen3-Next eager mode to v0.11.2 (#4477)

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

Adapted Qwen3-Next eager mode to `v0.11.2`.


- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

Signed-off-by: drslark <slarksblood@qq.com>
This commit is contained in:
drslark
2025-11-27 17:44:59 +08:00
committed by GitHub
parent b220de33e8
commit 1cae3e4a49
3 changed files with 16 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_ascend/models/qwen3_next.py
View File

@@ -675,7 +675,7 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
initial_state[~has_initial_state, ...] = 0
batch_size = initial_state.shape[0]
core_attn_out = []
temp_core_attn_out = []
last_recurrent_state = []
for b_idx in range(batch_size):
@@ -702,18 +702,18 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
use_qk_l2norm_in_kernel=True,
)
core_attn_out.append(cur_core_attn_out_non_spec)
temp_core_attn_out.append(cur_core_attn_out_non_spec)
last_recurrent_state.append(cur_last_recurrent_state)
tar_dtype = core_attn_out[0].dtype
tar_device = core_attn_out[0].device
tar_shape = list(core_attn_out[0].shape)
tar_dtype = temp_core_attn_out[0].dtype
tar_device = temp_core_attn_out[0].device
tar_shape = list(temp_core_attn_out[0].shape)
tar_shape[1] = non_spec_query_start_loc[-1]
core_attn_out_non_spec = torch.empty(tar_shape,
dtype=tar_dtype,
device=tar_device)
for b_idx in range(batch_size):
cur_core_attn_out = core_attn_out[b_idx]
cur_core_attn_out = temp_core_attn_out[b_idx]
start, end = non_spec_query_start_loc[
b_idx], non_spec_query_start_loc[b_idx + 1]
core_attn_out_non_spec[:, start:end, ...] = cur_core_attn_out