[perf]: support dual-batch overlap(dbo) for deepseek (#941)
### What this PR does / why we need it? Based on the design of dual-batch overlap proposed by Deepseek team and also the implementation of fused moe in VLLM project, we implement the multi-stream(also known as dual-batch) overlap for deepseek+mla on Ascend NPU. We split the input batch of model into two microbatches and then overlap the comp/comm ops in attention and moe layers using two streams to improve the performance. Our approach can be easily extended when adding dispatch/combine communications for moe layer. Compared with the previously proposed [draft](https://github.com/vllm-project/vllm-ascend/pull/842), we use one stream for computation ops and the other for communication ops, separately. In out opinions, it is beneficial for arranging the order of executing different ops and thus avoiding the contention of computation/communication resources. ref: [overlap for llama](https://github.com/vllm-project/vllm/pull/15787/files) ref: [dbo in sglang](https://github.com/sgl-project/sglang/pull/4068/files#diff-b4937569fc71f6ad215181b633b2f89c7183a2b4ac39e41fc22635599a9be7de) ### Does this PR introduce _any_ user-facing change? Adding an env variable "VLLM_ENABLE_DBO". Users can enable dbo by setting "VLLM_ASCEND_ENABLE_DBO=1" See /examples/offline_dualbatch_overlap_npu.py for more info. ### How was this patch tested? This patch can be tested with vllm-0.9.0 using its online service with benchmark tests. We have decoupled the func of dbo from vllm and it should be able to run without any modification to the code of vllm(some modifications is better to implement in vllm though). Any advice/discussion is welcome. ### Performance Benchmark We have ran the benchmark_serving script of vllm to test the performance after using dual-batch overlap. `python -m vllm.entrypoints.openai.api_server \ --model=DeepSeek-R1-W8A8 \ --trust-remote-code \ --distributed-executor-backend=mp \ -tp=16 \ --port 8006 \ --max-num-seqs 390 \ --max-model-len 32768 \ --max-num-batched-tokens 65536 \ --block-size 128 \ --compilation_config 0 \ --gpu-memory-utilization 0.90 \ --disable-log-requests \ --additional-config '{"expert_tensor_parallel_size":1,"enable_inter_dp_scheduling":true,"init_torchair_graph_batch_sizes":true,"trace_recompiles":true,"ascend_scheduler_config":{},"enable_graph_mode":false}'` and run benchmark with the parameters of : `--dataset-name random --random-input-len 4096 --random-output-len 1 --num-prompts 200 --max-concurrency 8 --request-rate 5 --metric-percentiles 90` 1. test with the version using allgather+allreduce in Ascend 910B (tp16 ep16 + deepseek r1 w8a8) 2. test with the version using alltoall: prefill qps: 0.90 -> 1.01 Mean TTFT：8226->7432ms The overlap approach when using alltoall communication can be further optimized by overlapping micro-batch1's moe comp with micro-batch2's dispatch a2a comm --------- Signed-off-by: zhuohuan <zxdu1997@gmail.com>
This commit is contained in:
zxdukki
@@ -13,6 +13,9 @@ from vllm.model_executor.layers.linear import (LinearBase,
|
||||
|
||||
from vllm_ascend.ascend_config import get_ascend_config
|
||||
from vllm_ascend.attention.attention_v1 import AscendAttentionState
|
||||
from vllm_ascend.multistream.base import MSAttentionMetadataSplitConfig
|
||||
from vllm_ascend.multistream.context import get_multistream_comm_context
|
||||
from vllm_ascend.multistream.ms_split import model_input_split_v1_mla_attn
|
||||
from vllm_ascend.ops.attention import vanilla_chunked_prefill_mla
|
||||
|
||||
if TYPE_CHECKING:
|
||||
@@ -117,6 +120,7 @@ class AscendMLAMetadata:
|
||||
|
||||
with_prefill_across_dp: bool = False
|
||||
|
||||
query_lens: Optional[list[int]] = None
|
||||
# The dimension of the attention heads
|
||||
head_dim: Optional[int] = None
|
||||
attn_mask: torch.Tensor = None
|
||||
@@ -135,6 +139,17 @@ class AscendMLAMetadata:
|
||||
# f"Only {supported_head_sizes} are supported for head_dim,",
|
||||
# f"received {self.head_dim}.")
|
||||
|
||||
def split_metadata_for_multistream(
|
||||
self,
|
||||
ms_split_config: MSAttentionMetadataSplitConfig,
|
||||
) -> list["AscendMLAMetadata"]:
|
||||
"""Split metadata for multi-stream with AscendMLAMetadata"""
|
||||
return model_input_split_v1_mla_attn(
|
||||
ms_split_config=ms_split_config,
|
||||
attn_metadata=self,
|
||||
_metadata_cls=AscendMLAMetadata,
|
||||
)
|
||||
|
||||
|
||||
M = TypeVar("M", bound=AscendMLAMetadata)
|
||||
|
||||
@@ -386,6 +401,7 @@ class AscendMLAMetadataBuilder:
|
||||
|
||||
return self.metadata_cls( # type: ignore
|
||||
num_actual_tokens=num_actual_tokens,
|
||||
query_lens=query_lens.tolist(),
|
||||
slot_mapping=slot_mapping,
|
||||
head_dim=self.runner.model_config.get_head_size(),
|
||||
num_decodes=self._num_decodes,
|
||||
@@ -585,7 +601,15 @@ class AscendMLAImpl(MLAAttentionImpl):
|
||||
)
|
||||
attn_output = attn_output.reshape(
|
||||
[num_tokens, self.num_heads * self.v_head_dim])
|
||||
return self.o_proj(attn_output)[0]
|
||||
|
||||
current_ms_metadata = get_multistream_comm_context()
|
||||
if current_ms_metadata is None:
|
||||
return self.o_proj(attn_output)[0]
|
||||
else:
|
||||
current_ms_metadata.before_comm_event.record()
|
||||
with torch.npu.stream(current_ms_metadata.comm_stream):
|
||||
current_ms_metadata.before_comm_event.wait()
|
||||
return self.o_proj(attn_output)[0]
|
||||
|
||||
def exec_kv(
|
||||
self,
|
||||
@@ -685,7 +709,14 @@ class AscendMLAImpl(MLAAttentionImpl):
|
||||
context_lens=attn_metadata.decode.seq_lens, # type:ignore
|
||||
mla_vheadsize=self.kv_lora_rank,
|
||||
out=attn_output)
|
||||
return self._v_up_proj_and_o_proj(attn_output)
|
||||
current_ms_metadata = get_multistream_comm_context()
|
||||
if current_ms_metadata is None:
|
||||
return self._v_up_proj_and_o_proj(attn_output)
|
||||
else:
|
||||
current_ms_metadata.before_comm_event.record()
|
||||
with torch.npu.stream(current_ms_metadata.comm_stream):
|
||||
current_ms_metadata.before_comm_event.wait()
|
||||
return self._v_up_proj_and_o_proj(attn_output)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -811,16 +842,38 @@ class AscendMLAImpl(MLAAttentionImpl):
|
||||
key_cache=kv_cache,
|
||||
slot_indices=attn_metadata.slot_mapping.flatten())
|
||||
if has_prefill:
|
||||
output[num_decode_tokens:] = self._forward_prefill(
|
||||
prefill_q, prefill_k_c_normed, prefill_k_pe, kv_cache,
|
||||
attn_metadata)
|
||||
# FIX: aicore move should be also placed on the comm stream in dbo,
|
||||
# otherwise it may affect the accuracy
|
||||
# TODO: use an elegant way to overlap
|
||||
output_prefill = self._forward_prefill(prefill_q,
|
||||
prefill_k_c_normed,
|
||||
prefill_k_pe, kv_cache,
|
||||
attn_metadata)
|
||||
current_ms_metadata = get_multistream_comm_context()
|
||||
if current_ms_metadata is not None:
|
||||
with torch.npu.stream(current_ms_metadata.comm_stream):
|
||||
output[num_decode_tokens:] = output_prefill
|
||||
current_ms_metadata.after_comm_event.record()
|
||||
else:
|
||||
output[num_decode_tokens:] = output_prefill
|
||||
|
||||
if has_decode:
|
||||
if self.running_in_graph:
|
||||
return self._forward_decode(decode_ql_nope, decode_q_pe,
|
||||
decode_k_nope, decode_k_pe,
|
||||
kv_cache, attn_metadata)
|
||||
else:
|
||||
output[:num_decode_tokens] = self._forward_decode(
|
||||
decode_ql_nope, decode_q_pe, decode_k_nope, decode_k_pe,
|
||||
kv_cache, attn_metadata)
|
||||
output_decode = self._forward_decode(decode_ql_nope,
|
||||
decode_q_pe,
|
||||
decode_k_nope,
|
||||
decode_k_pe, kv_cache,
|
||||
attn_metadata)
|
||||
current_ms_metadata = get_multistream_comm_context()
|
||||
if current_ms_metadata is not None:
|
||||
with torch.npu.stream(current_ms_metadata.comm_stream):
|
||||
output[:num_decode_tokens] = output_decode
|
||||
current_ms_metadata.after_comm_event.record()
|
||||
else:
|
||||
output[:num_decode_tokens] = output_decode
|
||||
|
||||
return output_padded
|
||||
|
||||
Reference in New Issue
Block a user