iHore/vLLM-4x-RTX-3090
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: Adjust start script to avoid crashing on high pressure KV cache

Browse Source 
Stress-tested using 4 concurrent requests with 240K tokens each:
```
(APIServer pid=1) INFO 10-15 06:21:25 [loggers.py:315] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3.0 tokens/s, Running: 4 reqs, Waiting: 1 reqs, GPU KV cache usage: 98.8%, Prefix cache hit rate: 0.0%
(APIServer pid=1) INFO 10-15 06:21:26 [loggers.py:315] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 17.0 tokens/s, Running: 4 reqs, Waiting: 1 reqs, GPU KV cache usage: 99.2%, Prefix cache hit rate: 0.0%
(APIServer pid=1) INFO 10-15 06:21:27 [loggers.py:315] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8.0 tokens/s, Running: 4 reqs, Waiting: 1 reqs, GPU KV cache usage: 99.6%, Prefix cache hit rate: 0.0%
(APIServer pid=1) INFO 10-15 06:21:28 [loggers.py:315] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 20.9 tokens/s, Running: 4 reqs, Waiting: 1 reqs, GPU KV cache usage: 100.0%, Prefix cache hit rate: 0.0%
```
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
MOVZX
2026-10-15 06:24:25 +00:00
parent 777e910dd0
commit eb8c302bb3
1 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
start.sh
+5 -4
View File
@@ -258,18 +258,19 @@ exec ${topo_launch_prefix:+$topo_launch_prefix} "$VENV/bin/vllm" serve "$CKPT" \
--host "${HOST:-127.0.0.1}" --port "${PORT:-8000}" \ --host "${HOST:-127.0.0.1}" --port "${PORT:-8000}" \
--tensor-parallel-size 2 --pipeline-parallel-size 2 --enable-expert-parallel \ --tensor-parallel-size 2 --pipeline-parallel-size 2 --enable-expert-parallel \
--speculative-config '{"method":"mtp","num_speculative_tokens":3,"draft_sample_method":"probabilistic"}' \ --speculative-config '{"method":"mtp","num_speculative_tokens":3,"draft_sample_method":"probabilistic"}' \
--max-model-len 262144 --max-num-seqs 2 --max-num-batched-tokens 4096 \ --max-model-len 262144 --max-num-seqs 4 --max-num-batched-tokens 2048 \
--gpu-memory-utilization 0.96 --kv-cache-memory 4100000000 --kv-cache-dtype fp8_e4m3 \ --gpu-memory-utilization 0.96 --kv-cache-memory 4100000000 --kv-cache-dtype fp8_e4m3 \
--limit-mm-per-prompt '{"image":42,"video":0}' --mm-processor-cache-gb 0 \ --limit-mm-per-prompt '{"image":42,"video":0}' --mm-processor-cache-gb 0 \
--mm-processor-kwargs '{"size":{"longest_edge":4194304,"shortest_edge":65536}}' \ --mm-processor-kwargs '{"size":{"longest_edge":4194304,"shortest_edge":65536}}' \
--compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[1,2,3,4,6,8,12,16,24,32]}' \ --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[1,2,4,8,12,16]}' \
--enable-prefix-caching --enable-prompt-tokens-details \ --enable-prefix-caching --enable-prompt-tokens-details \
--no-enable-flashinfer-autotune \ --no-enable-flashinfer-autotune \
--enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 \ --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 \
--override-generation-config '{"temperature": 1.0, "top_p": 0.95, "top_k": 20, "min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0}' \ --override-generation-config '{"temperature": 1.0, "top_p": 0.95, "top_k": 20, "min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0}' \
--default-chat-template-kwargs '{"enable_thinking": true, "reasoning_effort": "low"}' \ --default-chat-template-kwargs '{"enable_thinking": true, "reasoning_effort": "xhigh"}' \
--shutdown-timeout 10 \ --shutdown-timeout 10 \
--block-size 4096 --use-replayssm \ --block-size 2048 --use-replayssm \
--prefix-match-unit 64 \ --prefix-match-unit 64 \
--safetensors-load-strategy prefetch \ --safetensors-load-strategy prefetch \
--disable-uvicorn-access-log \
"$@" "$@"