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

[Model] GLM5 adaptation (#6642)

Browse Source 
### What this PR does / why we need it?
GLM5 adaptation
1. use torch_npu.npu_lightning_indexer for GLM5
2. forbid eagle proposer when fullgraph mode is enabled because of bugs
3. add quatization config for GLM5
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
- vLLM main:
978a37c823

---------

Signed-off-by: yydyzr <liuyuncong1@huawei.com>
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
Co-authored-by: shenchuxiaofugui <1311027364@qq.com>
This commit is contained in:
yydyzr
2026-02-11 22:22:22 +08:00
committed by GitHub
parent 140fcaffc3
commit ff3a50d011
17 changed files with 77 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
vllm_ascend/attention/sfa_v1.py
View File

@@ -431,6 +431,11 @@ class AscendSFAImpl(MLAAttentionImpl):
self.weights_proj = self.indexer.weights_proj
self.k_norm = self.indexer.k_norm
self.cp_size = 1
self.is_rope_neox_style = True
self.use_torch_npu_lightning_indexer = False
if self.vllm_config.model_config.hf_config.model_type in ["glm_moe_dsa"]:
self.is_rope_neox_style = False
self.use_torch_npu_lightning_indexer = True
self.enable_dsa_cp = enable_dsa_cp()
self.enable_dsa_cp_prefill_only = enable_dsa_cp_with_layer_shard()
@@ -973,7 +978,9 @@ class AscendSFAImpl(MLAAttentionImpl):
cos = cos.view(-1, self.qk_rope_head_dim)
sin = sin.view(-1, self.qk_rope_head_dim)
q, k = rope_forward_triton(q, k, cos, sin, rope_dim=self.qk_rope_head_dim, is_neox_style=True)
q, k = rope_forward_triton(
q, k, cos, sin, rope_dim=self.qk_rope_head_dim, is_neox_style=self.is_rope_neox_style
)
else:
k_pe, k_nope = torch.split(k, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1)
@@ -1036,18 +1043,35 @@ class AscendSFAImpl(MLAAttentionImpl):
key = self.gather_kv_cross_cp(key, attn_metadata.sfa_cp_metadata.valid_block_ids)
block_table = attn_metadata.sfa_cp_metadata.block_table_cp
topk_indices = torch.ops._C_ascend.npu_lightning_indexer(
query=q,
key=key,
weights=weights,
actual_seq_lengths_query=actual_seq_lengths_query,
actual_seq_lengths_key=actual_seq_lengths_key,
block_table=block_table,
layout_query="TND",
layout_key="PA_BSND",
sparse_count=2048,
sparse_mode=3,
)
# DSV3.2 currently has graph compilation issues when using torch_npu.npu.lightning_indexer.
# So two branches are maintained temporarily.
# TODO: torch.ops._C_ascend.npu_lightning_indexer needs to be removed.
if self.use_torch_npu_lightning_indexer:
topk_indices, _ = torch_npu.npu_lightning_indexer(
query=q,
key=key,
weights=weights,
actual_seq_lengths_query=actual_seq_lengths_query,
actual_seq_lengths_key=actual_seq_lengths_key,
block_table=block_table,
layout_query="TND",
layout_key="PA_BSND",
sparse_count=2048,
sparse_mode=3,
)
else:
topk_indices = torch.ops._C_ascend.npu_lightning_indexer(
query=q,
key=key,
weights=weights,
actual_seq_lengths_query=actual_seq_lengths_query,
actual_seq_lengths_key=actual_seq_lengths_key,
block_table=block_table,
layout_query="TND",
layout_key="PA_BSND",
sparse_count=2048,
sparse_mode=3,
)
return topk_indices
def _init_o_proj_tp_full_params(self):