0664c6e67a
### What this PR does / why we need it?
#### Documentation Improvements
New Configuration: Added the layer_sharding parameter to the
DeepSeek-V3.2-W8A8 deployment tutorial. This guides users to include
`["q_b_proj", "o_proj"]` in their prefill node setup for better resource
utilization.
#### CI and Testing Updates
Test Config Update: Updated the multi-node E2E test configuration file:
tests/e2e/nightly/multi_node/config/DeepSeek-V3_2-W8A8-A3-dual-nodes.yaml.
including disable `FLASHCOMM` and enable `FULL_DECODE_ONLY` and update
performance baseline.
### Does this PR introduce any user-facing change?
Yes. The documentation now recommends a more optimized startup command
for DeepSeek-V3.2-W8A8. Users following the updated tutorial will see
improved performance in multi-node PD disaggregation environments.
### How was this patch tested?
CI Validation: The updated E2E test configuration has been verified
through the nightly CI pipeline.
Environment: * vLLM version: v0.13.0
Base Commit:
[11b6af5](11b6af5280)
Hardware: Ascend A3/A2 multi-node cluster.
---------
Signed-off-by: guozr <guozr1997@hotmail.com>
Co-authored-by: guozr <guozr1997@hotmail.com>
88 lines
2.7 KiB
YAML
88 lines
2.7 KiB
YAML
test_name: "test DeepSeek-V3.2-W8A8 on A3"
|
|
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
|
|
num_nodes: 2
|
|
npu_per_node: 16
|
|
env_common:
|
|
HCCL_OP_EXPANSION_MODE: "AIV"
|
|
|
|
VLLM_USE_MODELSCOPE: true
|
|
HCCL_BUFFSIZE: 1024
|
|
SERVER_PORT: 8080
|
|
OMP_PROC_BIND: false
|
|
OMP_NUM_THREADS: 1
|
|
VLLM_ASCEND_ENABLE_MLAPO: 1
|
|
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
|
VLLM_ASCEND_ENABLE_FLASHCOMM1: 0
|
|
ASCEND_A3_EBA_ENABLE: 1
|
|
|
|
|
|
deployment:
|
|
-
|
|
server_cmd: >
|
|
vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
|
|
--host 0.0.0.0
|
|
--port $SERVER_PORT
|
|
--data-parallel-size 4
|
|
--data-parallel-size-local 2
|
|
--data-parallel-address $LOCAL_IP
|
|
--data-parallel-rpc-port 13399
|
|
--tensor-parallel-size 8
|
|
--quantization ascend
|
|
--seed 1024
|
|
--enable-expert-parallel
|
|
--max-num-seqs 16
|
|
--max-model-len 8192
|
|
--max-num-batched-tokens 4096
|
|
--no-enable-prefix-caching
|
|
--gpu-memory-utilization 0.85
|
|
--trust-remote-code
|
|
--speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
|
|
--compilation-config '{"cudagraph_capture_sizes": [3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 36, 39, 42, 45, 48], "cudagraph_mode": "FULL_DECODE_ONLY"}'
|
|
--tokenizer-mode deepseek_v32
|
|
--reasoning-parser deepseek_v3
|
|
|
|
-
|
|
server_cmd: >
|
|
vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
|
|
--headless
|
|
--data-parallel-size 4
|
|
--data-parallel-rpc-port 13399
|
|
--data-parallel-size-local 2
|
|
--data-parallel-start-rank 2
|
|
--data-parallel-address $MASTER_IP
|
|
--tensor-parallel-size 8
|
|
--quantization ascend
|
|
--seed 1024
|
|
--enable-expert-parallel
|
|
--max-num-seqs 16
|
|
--max-model-len 8192
|
|
--max-num-batched-tokens 4096
|
|
--no-enable-prefix-caching
|
|
--gpu-memory-utilization 0.85
|
|
--trust-remote-code
|
|
--speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
|
|
--compilation-config '{"cudagraph_capture_sizes": [3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 36, 39, 42, 45, 48], "cudagraph_mode": "FULL_DECODE_ONLY"}'
|
|
--tokenizer-mode deepseek_v32
|
|
--reasoning-parser deepseek_v3
|
|
benchmarks:
|
|
perf:
|
|
case_type: performance
|
|
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
|
request_conf: vllm_api_stream_chat
|
|
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
|
num_prompts: 512
|
|
max_out_len: 3000
|
|
batch_size: 512
|
|
request_rate: 11.2
|
|
baseline: 905.6805
|
|
threshold: 0.97
|
|
acc:
|
|
case_type: accuracy
|
|
dataset_path: vllm-ascend/gsm8k-lite
|
|
request_conf: vllm_api_general_chat
|
|
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
|
|
max_out_len: 4096
|
|
batch_size: 64
|
|
baseline: 95
|
|
threshold: 5
|