99a22f05b8
The previous params caused all 4 DP replicas to bind to the same 4 dies (die 0-3), leaving 12 dies idle. Root cause was a combination of missing official params + extra non-official params that interfered with DP worker device placement. Verified: with aligned params, TP=4 DP=4 correctly distributes 16 workers across all 16 dies (8 cards x 2 dies), each at ~57 GB HBM (91% util). Changes (align to docs.vllm.ai A3 tutorial): - Add --max-num-batched-tokens 10240 (was missing; affects DP scheduling) - Add --api-server-count 1 (was missing; without it vllm spawns N API servers for N DP ranks, disturbing device assignment) - Remove --kv-cache-dtype fp8 (official uses default bfloat16) - Remove --trust-remote-code (official doesn't use it for DSV4) - Remove enable_dsa_cp from additional-config (official doesn't have it) - max-model-len: per-TP caps (32768/65536/131072) -> 1048576 for all TPs (official uses full 1M; the caps were over-cautious) - max-num-seqs: per-TP (128/256/256) -> 64 for all (official value)
151 lines
7.7 KiB
Bash
151 lines
7.7 KiB
Bash
# TP×DP matrix experiment for DeepSeek-V4-Flash on Ascend 910C (8 NPUs / 16 dies)
|
||
# using vLLM-Ascend.
|
||
# Tests vLLM with three parallel configurations:
|
||
# TP=2, DP=4 -> 2 dies per replica, 4 replicas
|
||
# TP=4, DP=2 -> 4 dies per replica, 2 replicas
|
||
# TP=8, DP=1 -> 8 dies, no data parallelism
|
||
#
|
||
# Platform: ascend_910c (see platforms/ascend_910c.env).
|
||
# Host: 910c.1 / NPU-NODE61, openEuler 22.03 SP4 aarch64, driver 25.5.2, CANN 9.0.0.
|
||
#
|
||
# DeepSeek-V4-Flash w8a8-mtp weights are present at /mnt/models/DeepSeek-V4-Flash-w8a8-mtp
|
||
# (downloaded from ModelScope Eco-Tech/DeepSeek-V4-Flash-w8a8-mtp, ~280 GiB, 70 shards).
|
||
|
||
EXPERIMENT="dsv4_910c_vllm_tp_dp_matrix"
|
||
MODEL_NAME="DeepSeek-V4-Flash"
|
||
# Real DSV4-Flash w8a8-mtp directory (verified present & chown'd to shishi).
|
||
MODEL_PATH="${MODEL_PATH:-/mnt/models/DeepSeek-V4-Flash-w8a8-mtp}"
|
||
SERVED_MODEL_NAME="dsv4"
|
||
|
||
VLLM_PORT="${VLLM_PORT:-30052}"
|
||
|
||
# Dedicated container name so this experiment never touches other 910c runs.
|
||
CONTAINER_NAME="vllm-ascend-dsv4-910c"
|
||
|
||
# Python interpreter for the benchmark client inside the vllm-ascend container.
|
||
CONTAINER_PYTHON="/usr/local/python3.12.13/bin/python3"
|
||
|
||
# vllm-ascend image. Use the general v0.23 A3 image for DSV4 (the GLM5.2-tuned
|
||
# variant carries GLM-specific patches and is NOT compatible with DSV4).
|
||
# Verified present locally: quay.io/ascend/vllm-ascend:v0.23.0rc1-a3-openeuler
|
||
USE_DOCKER="${USE_DOCKER:-1}"
|
||
DOCKER_IMAGE="${DOCKER_IMAGE:-local/vllm-ascend:0.23-a3-dsv4-sglang}"
|
||
|
||
# Benchmark client Docker image. vLLM's image does not include sglang.bench_serving;
|
||
# reuse the vllm-ascend container itself for the client via `docker exec` (see
|
||
# run_adaptive_concurrency_add16.sh), so this is only used if USE_DOCKER_CLIENT=1
|
||
# with an external sglang image. Default off on 910c.
|
||
DOCKER_CLIENT_IMAGE="${DOCKER_CLIENT_IMAGE:-lmsysorg/sglang:latest}"
|
||
USE_DOCKER_CLIENT="${USE_DOCKER_CLIENT:-0}"
|
||
|
||
# Device selection. We mount all 8 cards (16 dies) via --device /dev/davinci0..15
|
||
# in start_vllm_docker.sh. ASCEND_VISIBLE_DEVICES is kept for parity with the
|
||
# shared library but the explicit --device flags are the authoritative path on
|
||
# this host (Ascend Docker Runtime injection was unreliable here).
|
||
export ASCEND_VISIBLE_DEVICES="${ASCEND_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
|
||
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
|
||
|
||
# Runtime working directory for logs, pid files, and tmp.
|
||
RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}"
|
||
|
||
# Parallel configurations to test. Format: "TP DP"
|
||
# A3 910C has 16 dies (8 cards x 2 dies/card). TP/DP address dies, so TP*DP=16
|
||
# uses all dies. DSV4-Flash w8a8 weights ~280 GiB total, sharded across TP dies:
|
||
# TP=4 DP=4 -> 4 dies/replica x 4 replicas (39 GiB/die, tight KV cache)
|
||
# TP=8 DP=2 -> 8 dies/replica x 2 replicas (35 GiB/die, balanced)
|
||
# TP=16 DP=1 -> 16 dies/replica x 1 replica (17.5 GiB/die, max KV cache)
|
||
# Note: A3 TP=4 == H20 TP=4 per-card-equivalent (A3 has 2 dies/card).
|
||
# TP=2, DP=4 -> 4 dies/replica x 4 replicas (smallest TP, most replicas)
|
||
# TP=4, DP=2 -> 4 dies/replica x 2 replicas
|
||
# TP=8, DP=1 -> 8 dies/replica x 1 replica (largest TP, max weight sharding)
|
||
if [[ -n "${PARALLEL_CONFIGS_STR:-}" ]]; then
|
||
declare -a PARALLEL_CONFIGS=()
|
||
for pair in $PARALLEL_CONFIGS_STR; do
|
||
PARALLEL_CONFIGS+=("${pair//,/ }")
|
||
done
|
||
else
|
||
declare -a PARALLEL_CONFIGS=(
|
||
"4 4"
|
||
"8 2"
|
||
"16 1"
|
||
)
|
||
fi
|
||
|
||
# vLLM-Ascend server settings for DSV4-Flash (w8a8-mtp).
|
||
# Notes:
|
||
# - KV cache dtype fp8 is supported on 910C; fall back to fp16 if the image rejects it.
|
||
# - block-size 128 matches Ascend page semantics (910C favors 128).
|
||
# - MAX_MODEL_LEN: DSV4-Flash supports up to 1M context; cap at 131072 for the
|
||
# matrix sweep (extend to 1M via matrix.json once TP=8 is verified).
|
||
# - gpu-memory-utilization maps to NPU HBM fraction on vllm-ascend (0.9).
|
||
GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.9}"
|
||
KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-}"
|
||
BLOCK_SIZE="${BLOCK_SIZE:-128}"
|
||
MAX_MODEL_LEN="${MAX_MODEL_LEN:-1048576}"
|
||
MAX_NUM_SEQS="${MAX_NUM_SEQS:-64}"
|
||
# Max tokens per batch (official A3 tutorial: 10240).
|
||
MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-10240}"
|
||
# API server count (official: 1).
|
||
API_SERVER_COUNT="${API_SERVER_COUNT:-1}"
|
||
|
||
# Per-TP parameter overrides (applied in start_vllm_docker.sh via case $TP).
|
||
# DSV4-Flash w8a8 weight per die = ~280GiB / TP. Each die has 64GB HBM.
|
||
# TP=4: ~39 GiB/die weights -> 25 GiB for KV cache; cap context to avoid OOM
|
||
# TP=8: ~35 GiB/die weights -> 29 GiB for KV cache; balanced
|
||
# TP=16: ~17.5 GiB/die weights -> 46 GiB for KV cache; full context
|
||
TP4_GPU_MEMORY_UTILIZATION="${TP4_GPU_MEMORY_UTILIZATION:-0.9}"
|
||
TP4_MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-1048576}"
|
||
TP4_MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-64}"
|
||
TP8_GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.9}"
|
||
TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-1048576}"
|
||
TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-64}"
|
||
TP16_GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.9}"
|
||
TP16_MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-1048576}"
|
||
TP16_MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-64}"
|
||
|
||
# vLLM-Ascend-specific launch flags injected by start_vllm_docker.sh.
|
||
# attention backend for 910C: use the fused/atb attention path. Adjust per image.
|
||
VLLM_ASCEND_ATTENTION_BACKEND="${VLLM_ASCEND_ATTENTION_BACKEND:-atb}"
|
||
|
||
# DSV4-Flash-specific server args (passed through to vllm serve in
|
||
# start_vllm_docker.sh). These are REQUIRED for DSV4 -- GLM-5.2 does not need them.
|
||
# --tokenizer-mode deepseek_v4 : DSV4 custom tokenizer
|
||
# --tool-call-parser deepseek_v4 : DSV4 tool-call parser
|
||
# --enable-auto-tool-choice : enable automatic tool choice
|
||
# --reasoning-parser deepseek_v4 : DSV4 reasoning parser
|
||
# --enable-expert-parallel : shard 256 experts across EP ranks
|
||
# --quantization ascend : use modelslim w8a8 quantization path
|
||
# --safetensors-load-strategy prefetch: prefetch shards for faster load
|
||
# --async-scheduling : overlap CPU scheduling with NPU compute
|
||
# MTP speculative decoding via SPEC_CONFIG
|
||
DSV4_TOKENIZER_MODE="${DSV4_TOKENIZER_MODE:-deepseek_v4}"
|
||
DSV4_TOOL_CALL_PARSER="${DSV4_TOOL_CALL_PARSER:-deepseek_v4}"
|
||
DSV4_REASONING_PARSER="${DSV4_REASONING_PARSER:-deepseek_v4}"
|
||
DSV4_QUANTIZATION="${DSV4_QUANTIZATION:-ascend}"
|
||
DSV4_SAFETENSORS_LOAD_STRATEGY="${DSV4_SAFETENSORS_LOAD_STRATEGY:-prefetch}"
|
||
# MTP speculative config (1 speculative token). JSON string, kept single-quoted
|
||
# in the launcher to avoid shell mangling.
|
||
DSV4_SPEC_CONFIG="${DSV4_SPEC_CONFIG:-{\"num_speculative_tokens\": 1, \"method\": \"mtp\", \"enforce_eager\": true}}"
|
||
DSV4_COMPILATION_CONFIG="${DSV4_COMPILATION_CONFIG:-{\"cudagraph_mode\": \"FULL_DECODE_ONLY\"}}"
|
||
DSV4_ADDITIONAL_CONFIG="${DSV4_ADDITIONAL_CONFIG:-{\"ascend_compilation_config\":{\"enable_npugraph_ex\":true,\"enable_static_kernel\":false},\"enable_cpu_binding\": true,\"multistream_overlap_shared_expert\":true}}"
|
||
DSV4_ENABLE_EXPERT_PARALLEL="${DSV4_ENABLE_EXPERT_PARALLEL:-1}"
|
||
DSV4_ENABLE_ASYNC_SCHEDULING="${DSV4_ENABLE_ASYNC_SCHEDULING:-1}"
|
||
|
||
# Model-loader extra config (multithread load, 128 threads) to speed up 280GiB load.
|
||
DSV4_MODEL_LOADER_EXTRA_CONFIG="${DSV4_MODEL_LOADER_EXTRA_CONFIG:-{\"enable_multithread_load\": \"true\", \"num_threads\": 128}}"
|
||
|
||
# Dataset used by sglang.bench_serving --dataset-name random.
|
||
DATASET_PATH="${DATASET_PATH:-${ROOT_DIR}/datasets/ShareGPT_V3_unfiltered_cleaned_split.json}"
|
||
|
||
# Matrix and concurrency rules are defined in matrix.json by default.
|
||
MATRIX_FILE="${MATRIX_FILE:-${SCRIPT_DIR:-.}/matrix.json}"
|
||
MATRIX_MODE="${MATRIX_MODE:-Y}"
|
||
|
||
export CONCURRENCY_SAMPLES="${CONCURRENCY_SAMPLES:-2}"
|
||
|
||
SCENARIO_TIMEOUT_S="${SCENARIO_TIMEOUT_S:-1800}"
|
||
GPU_MEM_SAMPLE_INTERVAL_S="${GPU_MEM_SAMPLE_INTERVAL_S:-1}"
|
||
|
||
DRY_RUN="${DRY_RUN:-0}"
|
||
GRID_LIMIT="${GRID_LIMIT:-0}"
|