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

52
vllm_ascend/worker/model_runner_v1.py
View File

@@ -84,10 +84,7 @@ from vllm_ascend.attention.utils import AscendCommonAttentionMetadata, using_pag
from vllm_ascend.compilation.acl_graph import (ACLGraphWrapper,
set_draft_graph_params,
set_graph_params,
update_attn_dcp_pcp_params,
update_attn_params,
update_mla_attn_dcp_pcp_params,
update_mla_attn_params)
update_full_graph_params)
# yapf: enable
from vllm_ascend.eplb.adaptor.vllm_adaptor import VllmEplbAdaptor
from vllm_ascend.eplb.core.eplb_device_transfer_loader import \
@@ -1142,26 +1139,9 @@ class NPUModelRunner(GPUModelRunner):
if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL \
and not self.use_sparse:
# TODO: maybe_padded_num_tokens will be removed, use num_input_tokens instead
if self.vllm_config.model_config.use_mla:
if self.pcp_size * self.dcp_size > 1:
# FIXME: Try using `auto_dispatch_capture=True`
update_mla_attn_dcp_pcp_params(self.update_stream,
forward_context,
maybe_padded_num_tokens)
else:
# FIXME: Try using `auto_dispatch_capture=True`
update_mla_attn_params(self.update_stream, forward_context,
maybe_padded_num_tokens,
self.speculative_config)
else:
if self.pcp_size * self.dcp_size > 1:
update_attn_dcp_pcp_params(self.update_stream,
forward_context,
maybe_padded_num_tokens)
else:
update_attn_params(self.update_stream, forward_context,
maybe_padded_num_tokens,
self.vllm_config)
update_full_graph_params(self.attn_backend, self.update_stream, forward_context,
maybe_padded_num_tokens, self.vllm_config,
self.vllm_config.speculative_config)
if get_forward_context().sp_enabled and not isinstance(
hidden_states, IntermediateTensors):
@@ -2038,25 +2018,9 @@ class NPUModelRunner(GPUModelRunner):
assert forward_context is not None
if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL and \
not forward_context.capturing and not self.use_sparse:
if self.vllm_config.model_config.use_mla:
# FIXME: Try using `auto_dispatch_capture=True`
if self.pcp_size * self.dcp_size > 1:
# FIXME: Try using `auto_dispatch_capture=True`
update_mla_attn_dcp_pcp_params(self.update_stream,
forward_context,
positions.shape[0])
else:
# FIXME: Try using `auto_dispatch_capture=True`
update_mla_attn_params(self.update_stream, forward_context,
num_tokens, self.speculative_config)
else:
if self.pcp_size * self.dcp_size > 1:
update_attn_dcp_pcp_params(self.update_stream,
forward_context,
positions.shape[0])
else:
update_attn_params(self.update_stream, forward_context,
num_tokens, self.vllm_config)
update_full_graph_params(self.attn_backend, self.update_stream, forward_context,
num_tokens, self.vllm_config,
self.speculative_config, positions.shape[0])
if self.use_aux_hidden_state_outputs:
hidden_states, _ = hidden_states
@@ -2899,7 +2863,7 @@ class NPUModelRunner(GPUModelRunner):
attn_layers = get_layers_from_vllm_config(self.vllm_config,
AttentionLayerBase)
# NOTE: Must process Attention/MLAAttention before MambaBase to maintain
# ordering expected by acl_graph.py's _update_attn_fia_params.
# ordering expected by graph parameter update logic in attention backends.
mamba_layers: dict[str, MambaBase] = {}
for layer_name, attn_module in attn_layers.items():
if isinstance(attn_module, Attention):