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

[Refactor] Unify full-graph parameter update logic (#6041)

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

**Refactor: Unify full-graph parameter update logic**

This PR consolidates the scattered full-graph parameter update logic
into a unified approach, improving code architecture and eliminating
duplication.

**Key improvements:**

1. **Unified interface**
- Create `update_full_graph_params` as the single entry point for all
full-graph updates
   - Replace multiple scattered update calls with one unified function
- Remove ~50 lines of duplicated if-else logic across
`model_runner_v1.py` and `eagle_proposer.py`

2. **Better architecture**
- Move update logic to respective Backend classes
(`AscendAttentionBackend`, `AscendMLABackend`)
   - Each Backend manages its own parameter update logic internally
   - Simplify caller code to just dispatch to the appropriate Backend

3. **Cleaner parameter handling**
   - Remove unnecessary `pcp_size` and `dcp_size` parameter passing
   - Get parallel configuration directly from distributed groups
   - Consistent with how other parts of the codebase obtain these values

**Why we need it:**
- **Maintainability**: Future changes only need to be made in one place
per Backend
- **Code quality**: Follows DRY principle and Single Responsibility
Principle
- **Readability**: Cleaner, more intuitive code structure

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

**No.** This is a pure refactoring with no functional changes - same
behavior, cleaner code.

### How was this patch tested?

- All existing unit tests pass with updated mocks
- No new tests needed (pure refactoring, no behavior changes)
- CI validates correctness

---

- vLLM version: v0.13.0

Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: drslark <slarksblood@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
This commit is contained in:
LICO67373
2026-01-24 20:12:57 +08:00
committed by GitHub
parent 8129c429ef
commit 8966a99710
10 changed files with 420 additions and 415 deletions
Download Patch File Download Diff File Expand all files Collapse all files

138
vllm_ascend/attention/attention_v1.py
View File

@@ -371,6 +371,144 @@ class AscendAttentionBackendImpl(AttentionImpl):
self.vllm_config.kv_transfer_config is not None and self.vllm_config.kv_transfer_config.is_kv_producer
)
@staticmethod
def update_graph_params(
update_stream,
forward_context,
num_tokens,
vllm_config,
speculative_config=None,
num_dcp_pcp_tokens=None,
):
if using_paged_attention(num_tokens, vllm_config):
# Paged Attention update logic
if forward_context.is_draft_model:
graph_params = get_draft_graph_params()
else:
graph_params = get_graph_params()
with torch.npu.stream(update_stream):
for key, param, handle, event in zip(
forward_context.attn_metadata,
graph_params.attn_params[num_tokens],
graph_params.handles[num_tokens],
graph_params.events[num_tokens],
):
(
query,
key_cache,
value_cache,
num_kv_heads,
num_heads,
scale,
block_table,
seq_lens,
output,
) = param
seq_lens = forward_context.attn_metadata[key].seq_lens
workspace = torch_npu._npu_paged_attention_get_workspace(
query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output,
)
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu._npu_paged_attention(
query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output,
workspace=workspace,
)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
else:
# FIA update logic
if forward_context.is_draft_model:
graph_params = get_draft_graph_params()
attn_metadata = forward_context.draft_attn_metadatas
attn_keys = list(attn_metadata[0].keys())
else:
graph_params = get_graph_params()
attn_metadata = forward_context.attn_metadata
attn_keys = list(attn_metadata.keys())
# For Qwen3-next, since the kv_cache_config has already categorized
# linear_attn and self_attn, the attn_metadata is first arranged with
# self_attn followed by linear_attn. Therefore, using zip directly
# filters out the update operations for linear_attn.
# TODO: We use a new variable `attn_keys` to ensure the loop count is
# correct after get by `zip` because of the new structure of the attn_metadata
# when running with the merged full eagle-graph. Should check it with Qwen3-next.
num_layers = len(attn_keys)
if num_layers == 0:
return
if forward_context.is_draft_model:
attn_keys = attn_keys * (len(graph_params.attn_params[num_tokens]) // num_layers)
attn_count = 0
with torch.npu.stream(update_stream):
for key, param, handle, event in zip(
attn_keys,
graph_params.attn_params[num_tokens],
graph_params.handles[num_tokens],
graph_params.events[num_tokens],
):
(
query,
key_cache,
value,
block_tables,
attn_mask,
block_size,
seq_lens,
query_start_loc,
num_kv_heads,
num_heads,
scale,
attn_output,
softmax_lse,
) = param
if forward_context.is_draft_model:
draft_step = attn_count // num_layers
seq_lens = attn_metadata[draft_step][key].seq_lens_list
actual_seq_lengths_q = attn_metadata[draft_step][key].actual_seq_lengths_q
attn_count = attn_count + 1
else:
seq_lens = attn_metadata[key].seq_lens_list
actual_seq_lengths_q = attn_metadata[key].actual_seq_lengths_q
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu.npu_fused_infer_attention_score.out(
query=query,
key=key_cache,
value=value,
block_table=block_tables,
atten_mask=attn_mask,
input_layout="TND",
block_size=block_size,
actual_seq_lengths=actual_seq_lengths_q,
actual_seq_lengths_kv=seq_lens,
num_key_value_heads=num_kv_heads,
num_heads=num_heads,
scale=scale,
sparse_mode=3,
workspace=graph_params.workspaces.get(num_tokens),
out=[attn_output, softmax_lse],
)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
def process_weights_after_loading(self, act_dtype: torch.dtype):
super().process_weights_after_loading(act_dtype)
if flashcomm2_oshard_manager.flashcomm2_oshard_enable():

73
vllm_ascend/attention/context_parallel/attention_cp.py
View File

@@ -276,6 +276,79 @@ class AscendAttentionCPImpl(AscendAttentionBackendImpl):
self.dcp_rank = get_decode_context_model_parallel_rank() if self.dcp_size > 1 else 0
self.dcp_group = get_dcp_group().device_group if self.dcp_size > 1 else None
@staticmethod
def update_graph_params(
update_stream,
forward_context,
num_tokens,
vllm_config,
speculative_config=None,
num_dcp_pcp_tokens=None,
):
graph_params = get_graph_params()
# FIXME: Behold! We are using a temporary hack here to update the args
# for each layer's attention op in the graph.
with torch.npu.stream(update_stream):
for key, param, handle, event in zip(
forward_context.attn_metadata,
graph_params.attn_params[num_tokens],
graph_params.handles[num_tokens],
graph_params.events[num_tokens],
):
(
q_nope,
k_nope,
value,
num_heads,
num_kv_heads,
scale,
block_table,
block_size,
actual_seq_lengths_kv,
actual_seq_lengths_q,
attn_output,
softmax_lse,
dcp_size,
pcp_rank,
dcp_rank,
) = param
attn_metadata = forward_context.attn_metadata[key]
actual_seq_lengths_kv = attn_metadata.decode_meta.num_computed_tokens_of_pcp_dcp[:, pcp_rank, dcp_rank]
pad_length = num_tokens - len(actual_seq_lengths_kv)
if pad_length > 0:
pad_tensor = np.zeros(pad_length, dtype=actual_seq_lengths_kv.dtype)
actual_seq_lengths_kv = np.concatenate([actual_seq_lengths_kv, pad_tensor])
actual_seq_lengths_q = attn_metadata.actual_seq_lengths_q
if dcp_size > 1:
num_heads = num_heads * dcp_size
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu.npu_fused_infer_attention_score.out(
q_nope,
k_nope,
value,
num_heads=num_heads,
num_key_value_heads=num_kv_heads,
input_layout="TND",
atten_mask=None,
scale=scale,
antiquant_mode=0,
antiquant_scale=None,
softmax_lse_flag=True,
block_table=block_table,
block_size=block_size,
actual_seq_lengths_kv=actual_seq_lengths_kv,
actual_seq_lengths=actual_seq_lengths_q,
workspace=graph_params.workspaces.get(num_tokens),
out=[attn_output, softmax_lse],
)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
def _attention_with_nomask_and_mask(
self,
q: torch.Tensor,

79
vllm_ascend/attention/context_parallel/mla_cp.py
View File

@@ -284,6 +284,85 @@ class AscendMlaCPImpl(AscendMLAImpl):
self.dcp_rank = get_decode_context_model_parallel_rank() if self.dcp_size > 1 else 0
self.dcp_group = get_dcp_group().device_group if self.dcp_size > 1 else None
@staticmethod
def update_graph_params(
update_stream,
forward_context,
num_tokens,
vllm_config=None,
speculative_config=None,
num_dcp_pcp_tokens=None,
):
if forward_context.is_draft_model:
graph_params = get_draft_graph_params()
else:
graph_params = get_graph_params()
# FIXME: Behold! We are using a temporary hack here to update the args
# for each layer's attention op in the graph.
with torch.npu.stream(update_stream):
for key, param, handle, event in zip(
forward_context.attn_metadata,
graph_params.attn_params[num_tokens],
graph_params.handles[num_tokens],
graph_params.events[num_tokens],
):
(
q_nope,
k_nope,
q_pe,
k_pe,
num_heads,
num_kv_heads,
input_layout,
spec_attn_mask,
sparse_mode,
scale,
block_table,
block_size,
actual_seq_lengths,
actual_seq_lengths_kv,
attn_output,
softmax_lse,
) = param
decode_meta = forward_context.attn_metadata[key].decode
seq_len = decode_meta.cp_seq_len
if isinstance(seq_len, torch.Tensor):
seq_len = seq_len.tolist()
actual_seq_lengths_kv = seq_len
pad_length = num_tokens - len(actual_seq_lengths_kv)
if pad_length > 0:
actual_seq_lengths_kv = actual_seq_lengths_kv + [0] * (num_tokens - len(actual_seq_lengths_kv))
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu.npu_fused_infer_attention_score.out(
q_nope,
k_nope,
k_nope,
query_rope=q_pe,
key_rope=k_pe,
num_heads=num_heads,
num_key_value_heads=num_kv_heads,
input_layout=input_layout,
atten_mask=spec_attn_mask,
sparse_mode=sparse_mode,
scale=scale,
antiquant_mode=0,
antiquant_scale=None,
softmax_lse_flag=True,
block_table=block_table,
block_size=block_size,
actual_seq_lengths_kv=actual_seq_lengths_kv,
actual_seq_lengths=actual_seq_lengths,
workspace=graph_params.workspaces.get(num_tokens),
out=[attn_output, softmax_lse],
)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
def get_num_actual_tokens(self, attn_metadata: M):
if self.pcp_size > 1:
return attn_metadata.num_actual_tokens_pcp_padded // self.pcp_size

82
vllm_ascend/attention/mla_v1.py
View File

@@ -745,6 +745,88 @@ class AscendMLAImpl(MLAAttentionImpl):
)
register_all_layers_to_shard_weight_series(self.layer_sharding_kwargs)
@staticmethod
def update_graph_params(
update_stream,
forward_context,
num_tokens,
vllm_config=None,
speculative_config=None,
num_dcp_pcp_tokens=None,
):
if forward_context.is_draft_model:
graph_params = get_draft_graph_params()
else:
graph_params = get_graph_params()
# FIXME: Behold! We are using a temporary hack here to update the args
# for each layer's attention op in the graph.
with torch.npu.stream(update_stream):
for key, param, handle, event in zip(
forward_context.attn_metadata,
graph_params.attn_params[num_tokens],
graph_params.handles[num_tokens],
graph_params.events[num_tokens],
):
(
q_nope,
k_nope,
q_pe,
k_pe,
num_heads,
num_kv_heads,
input_layout,
attn_mask,
sparse_mode,
scale,
block_table,
block_size,
seq_lens_list,
actual_seq_lengths,
attn_output,
softmax_lse,
) = param
seq_lens_list = forward_context.attn_metadata[key].decode.seq_lens_list
if speculative_config and speculative_config.method == "mtp" and not forward_context.is_draft_model:
actual_seq_lengths = forward_context.attn_metadata[key].decode.actual_seq_lengths_q
spec_multiple = speculative_config.num_speculative_tokens + 1
seq_lens_list = seq_lens_list + [0] * (num_tokens // spec_multiple - len(seq_lens_list))
actual_seq_lengths = [spec_multiple * (i + 1) for i in range(num_tokens // spec_multiple)]
elif forward_context.is_draft_model:
actual_seq_lengths = forward_context.attn_metadata[key].decode.actual_seq_lengths_q
block_table = forward_context.attn_metadata[key].decode.block_table
# TODO: This is a hack and should be fixed in the future.
if speculative_config.disable_padded_drafter_batch:
block_table = block_table[: len(actual_seq_lengths)]
seq_lens_list = seq_lens_list + [0] * (len(actual_seq_lengths) - len(seq_lens_list))
else:
seq_lens_list = seq_lens_list + [0] * (num_tokens - len(seq_lens_list))
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu.npu_fused_infer_attention_score.out(
q_nope,
k_nope,
k_nope,
query_rope=q_pe,
key_rope=k_pe,
num_heads=num_heads,
num_key_value_heads=num_kv_heads,
input_layout=input_layout,
atten_mask=attn_mask,
sparse_mode=sparse_mode,
scale=scale,
antiquant_mode=0,
antiquant_scale=None,
block_table=block_table,
block_size=block_size,
actual_seq_lengths_kv=seq_lens_list,
actual_seq_lengths=actual_seq_lengths,
workspace=graph_params.workspaces.get(num_tokens),
out=[attn_output, softmax_lse],
)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
def _v_up_proj(self, x):
# Convert from (N, B, L)/(N, B, 1, L) to (N, B, L)
x = x.view(self.num_heads, -1, self.kv_lora_rank)