Update README.md
Browse files
README.md
CHANGED
|
@@ -151,9 +151,108 @@ vllm serve \
|
|
| 151 |
--disable-uvicorn-access-log
|
| 152 |
```
|
| 153 |
|
| 154 |
-
Observed performance for this command on 8 × NVIDIA H200 141GB
|
| 155 |
(2026-08-05) is:
|
| 156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
| Item | Observed value |
|
| 158 |
|------|---------------:|
|
| 159 |
| Dynamic Cubic mode | A8 |
|
|
@@ -892,4 +991,4 @@ Both the code repository and the model weights are released under the [Kimi K3 L
|
|
| 892 |
|
| 893 |
## 8. Contact Us
|
| 894 |
|
| 895 |
-
If you have any questions, please reach out at [support@moonshot.ai](mailto:support@moonshot.ai).
|
|
|
|
| 151 |
--disable-uvicorn-access-log
|
| 152 |
```
|
| 153 |
|
| 154 |
+
Observed performance for this command on 8 × NVIDIA H200 141GB and 8 × NVIDIA H20 141GB
|
| 155 |
(2026-08-05) is:
|
| 156 |
|
| 157 |
+
### 【docker with vLLM Startup Command】
|
| 158 |
+
```bash
|
| 159 |
+
docker pull swr.cn-north-4.myhuaweicloud.com/desaysv/vllm-cubic:0.26.1-cubic.20260805
|
| 160 |
+
|
| 161 |
+
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
|
| 162 |
+
export VLLM_CUBIC_DYNAMIC_A8=1
|
| 163 |
+
export FLASHINFER_DISABLE_VERSION_CHECK=1
|
| 164 |
+
|
| 165 |
+
vllm serve \
|
| 166 |
+
__YOUR_PATH__/Kimi-K3-Cubic-2.5Bit \
|
| 167 |
+
--served-model-name Kimi-K3-Cubic-2.5Bit \
|
| 168 |
+
--trust-remote-code \
|
| 169 |
+
--quantization cubic \
|
| 170 |
+
--kv-cache-dtype fp8_q16 \
|
| 171 |
+
--gpu-memory-utilization 0.985 \
|
| 172 |
+
--tensor-parallel-size 8 \
|
| 173 |
+
--enable-expert-parallel \
|
| 174 |
+
--mm-encoder-tp-mode data \
|
| 175 |
+
--max-model-len auto \
|
| 176 |
+
--max-num-seqs 128 \
|
| 177 |
+
--max-num-batched-tokens 2048 \
|
| 178 |
+
--enable-prefix-caching \
|
| 179 |
+
--enable-auto-tool-choice \
|
| 180 |
+
--tool-call-parser kimi_k3 \
|
| 181 |
+
--reasoning-parser kimi_k3 \
|
| 182 |
+
--host 0.0.0.0 \
|
| 183 |
+
--port 8000 \
|
| 184 |
+
--disable-uvicorn-access-log
|
| 185 |
+
```
|
| 186 |
+
*skip deepgemm error ,its not effect start
|
| 187 |
+
|
| 188 |
+
|
| 189 |
+
### 【benchmark】10*8K/1K
|
| 190 |
+
|
| 191 |
+
| Item | Observed value |
|
| 192 |
+
|------|---------------:|
|
| 193 |
+
| Total requests | 10 |
|
| 194 |
+
| Successful requests | 10 |
|
| 195 |
+
| Failed requests | 0 |
|
| 196 |
+
| Unfinished requests | 0 |
|
| 197 |
+
| Average concurrency | 10 |
|
| 198 |
+
| Total duration | 209.75 s |
|
| 199 |
+
| Total throughput | 464.11 tokens/s |
|
| 200 |
+
| Input throughput | 684.34 tokens/s |
|
| 201 |
+
| Output throughput | 51.07 tokens/s |
|
| 202 |
+
| Average request latency | 204.23 s |
|
| 203 |
+
| Average TTFT | 69,607.26 ms |
|
| 204 |
+
| Average TPOT | 164.76 ms |
|
| 205 |
+
| Average ITL | 95.25 ms |
|
| 206 |
+
| TTFT (P50) | 63,297.56 ms |
|
| 207 |
+
| TTFT (P90) | 128,423.96 ms |
|
| 208 |
+
| TTFT (P99) | 128,423.96 ms |
|
| 209 |
+
| ITL (P50) | 81.85 ms |
|
| 210 |
+
| ITL (P90) | 178.56 ms |
|
| 211 |
+
| ITL (P99) | 178.56 ms |
|
| 212 |
+
| TPOT (P50) | 154.40 ms |
|
| 213 |
+
| TPOT (P90) | 206.35 ms |
|
| 214 |
+
| TPOT (P99) | 206.35 ms |
|
| 215 |
+
| Latency (P50) | 204.10 s |
|
| 216 |
+
| Latency (P90) | 208.45 s |
|
| 217 |
+
| Latency (P99) | 208.45 s |
|
| 218 |
+
| Input tokens per request | 8,088 |
|
| 219 |
+
| Output tokens per request | 1,000 |
|
| 220 |
+
|
| 221 |
+
### 【benchmark】10*64K/3000
|
| 222 |
+
|
| 223 |
+
| Item | Observed value |
|
| 224 |
+
|------|---------------:|
|
| 225 |
+
| Total requests | 10 |
|
| 226 |
+
| Successful requests | 10 |
|
| 227 |
+
| Failed requests | 0 |
|
| 228 |
+
| Unfinished requests | 0 |
|
| 229 |
+
| Average concurrency | 9 |
|
| 230 |
+
| Total duration | 1,461.29 s |
|
| 231 |
+
| Total throughput | 519.15 tokens/s |
|
| 232 |
+
| Input throughput | 513.63 tokens/s |
|
| 233 |
+
| Output throughput | 23.21 tokens/s |
|
| 234 |
+
| Average request latency | 1,387.96 s |
|
| 235 |
+
| Average TTFT | 882,028.37 ms |
|
| 236 |
+
| Average TPOT | 334.97 ms |
|
| 237 |
+
| Average ITL | 107.57 ms |
|
| 238 |
+
| TTFT (P50) | 604,938.80 ms |
|
| 239 |
+
| TTFT (P90) | 1,352,678.48 ms |
|
| 240 |
+
| TTFT (P99) | 1,352,678.48 ms |
|
| 241 |
+
| ITL (P50) | 61.83 ms |
|
| 242 |
+
| ITL (P90) | 313.74 ms |
|
| 243 |
+
| ITL (P99) | 313.74 ms |
|
| 244 |
+
| TPOT (P50) | 378.19 ms |
|
| 245 |
+
| TPOT (P90) | 465.73 ms |
|
| 246 |
+
| TPOT (P99) | 465.73 ms |
|
| 247 |
+
| Latency (P50) | 1,394.02 s |
|
| 248 |
+
| Latency (P90) | 1,460.10 s |
|
| 249 |
+
| Latency (P99) | 1,460.10 s |
|
| 250 |
+
| Input tokens per request | 64,088 |
|
| 251 |
+
| Output tokens per request | 3,000 |
|
| 252 |
+
|
| 253 |
+
|
| 254 |
+
|
| 255 |
+
|
| 256 |
| Item | Observed value |
|
| 257 |
|------|---------------:|
|
| 258 |
| Dynamic Cubic mode | A8 |
|
|
|
|
| 991 |
|
| 992 |
## 8. Contact Us
|
| 993 |
|
| 994 |
+
If you have any questions, please reach out at [support@moonshot.ai](mailto:support@moonshot.ai).
|