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

[Feat](sfa,dcp) support dcp for sfa (#6563)

Browse Source 
### What this PR does / why we need it?
This PR adds DCP support to the SFA backend.

Please note that due to operator constraints, the current implementation
has to all-gather the entire KV cache and modify the block table to
satisfy the operator input requirements. This results in significantly
increased communication overhead and peak memory usage. Therefore, this
is only a temporary workaround and will be refactored once the operator
provides proper support.

Additionally, because of the above limitations,
`cp_kv_cache_interleave_size` is currently required to be equal to
`block_size`. This restriction will also be removed after the refactor.

#### Test
accuracy test using DeepSeek-V3.2-Exp-W8A8 with dp2tp8dcp8

| dataset | version | metric | mode | vllm-api-general-stream |
|----- | ----- | ----- | ----- | -----|
| gsm8kdataset | - | accuracy | gen | 96.35 |

- vLLM version: v0.15.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0

---------

Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
This commit is contained in:
Qiu
2026-02-09 18:52:25 +08:00
committed by GitHub
parent 80e5812b39
commit cb7c419bc0
5 changed files with 190 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm_ascend/platform.py
View File

@@ -378,10 +378,11 @@ class NPUPlatform(Platform):
vllm_config.scheduler_config.enable_chunked_prefill = True
vllm_config.scheduler_config.SLO_limits_for_dynamic_batch = ascend_config.SLO_limits_for_dynamic_batch
cp_size = parallel_config.decode_context_parallel_size * parallel_config.prefill_context_parallel_size
if (
vllm_config.kv_transfer_config is not None
and cache_config.block_size != parallel_config.cp_kv_cache_interleave_size
and parallel_config.decode_context_parallel_size * parallel_config.prefill_context_parallel_size > 1
and cp_size > 1
):
raise AssertionError(
f"cp_kv_cache_interleave_size({parallel_config.cp_kv_cache_interleave_size}) "
@@ -389,6 +390,20 @@ class NPUPlatform(Platform):
"needs to be equal if use pcp or dcp > 1 in P/D disaggregate and kv pool scenario."
)
use_sparse = (
model_config is not None
and model_config.hf_text_config is not None
and hasattr(model_config.hf_text_config, "index_topk")
)
if use_sparse and cp_size > 1 and parallel_config.cp_kv_cache_interleave_size != cache_config.block_size:
logger.warning_once(
"The current SFA's PCP&DCP implementation requires"
f"cp_kv_cache_interleave_size({parallel_config.cp_kv_cache_interleave_size})"
f" == block_size({cache_config.block_size}). "
f"Override cp_kv_cache_interleave_size to {cache_config.block_size}."
)
vllm_config.parallel_config.cp_kv_cache_interleave_size = cache_config.block_size
if is_vl_model(vllm_config):
if bool(int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM", "0"))) or bool(
int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM1", "0"))