跳到主要内容

推理性能

模型性能数据

以下数据来自 SGLang-MUSA 0.5.10+20260803 在 M1000 AImodule 端侧环境的实测结果。均在性能模式 + compute_only 环境下测试,数值供参考。

测试环境

  • 端设备:M1000 AImodule,64G 统一内存
  • 系统版本:V1.4.1
  • MUSA 版本:5.1.1
  • 服务参数:max-total-tokens=16384
  • 文本测试并发:1
  • 文本测试输入 / 输出:无前缀场景为 128/100、250/100、500/100、1000/100、2000/100;带 prefix 场景为 500/100、1000/100、2000/100、4000/100、8000/100、16000/100,其中 PrefillLen=250

测试说明

测试指标说明如下:

  • Prefill(tps):Prefill 阶段吞吐,单位 token/s。
  • TTFT(ms):首 token 延迟,单位 ms。
  • Decode(tps):Decode 阶段吞吐,单位 token/s。
  • TPOT(ms):单个输出 token 平均耗时,单位 ms。
  • Profile:测试 profile,格式为 prefill_len:input_len
  • Vision(fps):视觉编码帧率,单位 frame/s。
  • Precision:模型量化精度。

性能复现流程

解压得到安装包目录:

tar -zxvf SGLang-MUSA-M1000-M1000_1.4.1_sglang.tar.gz
cd 20260830_M1000_1.4.1_sglang

修改 scripts/ci/musa/tilelang_workflow_config.json 中的模型路径为实际路径,比如:

"model_paths": [
"/home/dev/models/Qwen3-30B-A3B-GPTQ-Int4",
"/home/dev/models/Qwen3.5-35B-A3B-GPTQ-Int4"
]

生成工作流配置并加载环境变量,预编译 TileLang kernel;如果你已经在 快速开始 里执行过这一步,这里可以直接跳过:

scripts/ci/musa/generate_tilelang_workflow.sh
source scripts/ci/musa/setup_env.sh
scripts/ci/musa/precompile_tilelang_kernels.sh

分别启动服务并运行 benchmark。每组测试使用两个终端:启动 端运行服务,运行 benchmark 端执行对应测试。benchmark 通过 MODEL_KEY 选择对应模型配置。

Qwen3-30B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4 (W4A16))

启动:

scripts/ci/musa/start_qwen3_server.sh

运行 benchmark:

MODEL_KEY=qwen3 scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4 (W4A16))

启动:

scripts/ci/musa/start_qwen35_server.sh

运行 benchmark:

MODEL_KEY=qwen35 scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4 (W4A16)) 开启 MTP 投机解码

启动:

SGLANG_USE_TL_MOE_GROUPED_GEMM=0 QWEN35_EXTRA_SERVER_ARGS="--speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4" scripts/ci/musa/start_qwen35_server.sh

运行 benchmark:

MODEL_KEY=qwen35_mtp scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4 (W4A16)) 开启 DFLASH 投机解码

启动:

scripts/ci/musa/start_qwen35_dflash_server.sh

运行 benchmark:

MODEL_KEY=qwen35_dflash scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16))

启动:

scripts/ci/musa/start_qwen35_fq_server.sh

运行 benchmark:

MODEL_KEY=qwen35_fq scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5_FQ-35B-A3B(MoE GPTQ,GPTQ INT4 (W4A16)) 开启 MTP 投机解码

启动:

QWEN35_FQ_EXTRA_SERVER_ARGS="--speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4" scripts/ci/musa/start_qwen35_fq_server.sh

运行 benchmark:

MODEL_KEY=qwen35_fq_mtp scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5_FQ-35B-A3B(MoE GPTQ,GPTQ INT4 (W4A16)) 开启 DFLASH 投机解码

启动:

scripts/ci/musa/start_qwen35_fq_dflash_server.sh

运行 benchmark:

MODEL_KEY=qwen35_fq_dflash scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ, GPTQ INT4 (W4A16)) 开启 INT8 KV cache

启动:

scripts/ci/musa/start_qwen35_server.sh --kv-cache-dtype int8

运行 benchmark:

MODEL_KEY=qwen35 scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.6_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16))

启动:

scripts/ci/musa/start_qwen36_fq_server.sh

运行 benchmark:

MODEL_KEY=qwen36_fq scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3.6_FQ-35B-A3B(MoE GPTQ,GPTQ INT4 (W4A16)) 开启 DFLASH 投机解码

启动:

scripts/ci/musa/start_qwen36_fq_dflash_server.sh

运行 benchmark:

MODEL_KEY=qwen36_fq_dflash scripts/ci/musa/run_tilelang_benchmark.sh

Qwen3-VL-8B (VLM GPTQ,GPTQ INT4 (W4A16))

启动:

scripts/ci/musa/start_qwen3vl_server.sh

运行 benchmark:

MODEL_KEY=qwen3vl scripts/ci/musa/run_tilelang_benchmark.sh

参考性能数据

Qwen3-30B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4(W4A16))

ModelSizePrefillLen(tokens)Ctx(K)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)ProfilePrefix(tokens)MeasuredOutput(tokens)
Qwen330b_a3b128161112810064.6061981.22824.53140.765128:1280100
Qwen330b_a3b250161125010094.7312639.04123.96541.728250:2500100
Qwen330b_a3b5001611500100144.3823463.04323.31142.898500:5000100
Qwen330b_a3b100016111000100191.3715225.45323.02843.4261000:10000100
Qwen330b_a3b200016112000100181.55911015.68722.28544.8732000:20000100
Qwen330b_a3b250161150010077.033245.46823.59242.387250:500250100
Qwen330b_a3b2501611100010076.3033276.3922.87143.723250:1000750100
Qwen330b_a3b2501611200010071.8013481.84922.24544.954250:20001750100
Qwen330b_a3b2501611400010075.6123306.34920.89647.856250:40003750100
Qwen330b_a3b2501611800010062.9783969.64218.77953.251250:80007750100
Qwen330b_a3b25016111600010046.6545358.56715.66263.85250:1600015750100

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4(W4A16))

ModelSizePrefillLen(tokens)Ctx(K)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)ProfilePrefix(tokens)MeasuredOutput(tokens)
Qwen3.535b_a3b128161112810084.5991513.02614.81967.48128:1280100
Qwen3.535b_a3b250161125010085.5582921.98414.64768.271250:2500100
Qwen3.535b_a3b5001611500100141.8483524.89314.69968.032500:5000100
Qwen3.535b_a3b100016111000100213.3554687.0314.90567.0911000:10000100
Qwen3.535b_a3b200016112000100210.2239513.68914.5768.6322000:20000100
Qwen3.535b_a3b250161150010075.4973311.37314.57868.596250:500250100
Qwen3.535b_a3b2501611100010074.4513357.92814.54268.767250:1000750100
Qwen3.535b_a3b2501611200010073.7963387.72714.4969.011250:20001750100
Qwen3.535b_a3b2501611400010070.6943536.37714.55868.69250:40003750100
Qwen3.535b_a3b2501611800010075.1323327.47714.35369.671250:80007750100
Qwen3.535b_a3b25016111600010063.0763963.4913.82772.324250:1600015750100

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4(W4A16)) 开启 MTP 投机解码 (GSM8K测试集)

ModelSizePrefillLen(tokens)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)Prefix(tokens)Prompt Indexaccept rate((acclen-1)/(4-1))PredictionGround TruthCorrect
Qwen3.5+MTP35b_a3b86911869156184.474710.7620.7248.26060.74316464TRUE
Qwen3.5+MTP35b_a3b85511855116182.614682.1723.143.29070.8646260260TRUE
Qwen3.5+MTP35b_a3b88411884254186.864730.8221.5746.36080.7911160160TRUE
Qwen3.5+MTP35b_a3b91911919257191.124808.523.1943.11090.86854545TRUE
Qwen3.5+MTP35b_a3b87511875154185.024729.2622.9743.530100.8527460460TRUE
Qwen3.5+MTP35b_a3b87511875144185.874707.5323.7142.180110.8889366366TRUE
Qwen3.5+MTP35b_a3b87911879182186.184721.2823.6742.250120.898694694TRUE
Qwen3.5+MTP35b_a3b88211882179185.904744.5921.1347.320130.76541213FALSE
Qwen3.5+MTP35b_a3b87211872376185.324705.2723.3142.890140.88031818TRUE
Qwen3.5+MTP35b_a3b86711867223183.734718.7723.0443.410150.86026060TRUE
Qwen3.5+MTP35b_a3b90911909123190.234778.423.941.850160.899125125TRUE
Qwen3.5+MTP35b_a3b8681186861183.724724.6524.3241.110170.9167230230TRUE
Qwen3.5+MTP35b_a3b87211872148185.314705.5324.3741.040180.92315750057500TRUE
Qwen3.5+MTP35b_a3b8431184380180.894660.3221.28470190.763977TRUE
Qwen3.5+MTP35b_a3b88011880185186.344722.5121.9145.640200.802566TRUE
Qwen3.5+MTP35b_a3b87811878167185.794725.6922.744.060210.8441515TRUE
Qwen3.5+MTP35b_a3b86011860100183.314691.5222.7843.90220.84521414TRUE
Qwen3.5+MTP35b_a3b8701187091184.794707.9223.1543.190230.866777TRUE
Qwen3.5+MTP35b_a3b8551185571182.834676.4222.5644.320240.833388TRUE
Qwen3.5+MTP35b_a3b85411854171182.484680.0921.0847.430250.75642626TRUE

Qwen3.5-35B-A3B-GPTQ-Int4 (MoE GPTQ,GPTQ INT4(W4A16)) 开启 DFLASH 投机解码 (GSM8K测试集)

ModelSizePrefillLen(tokens)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)Prefix(tokens)PredictionGround Truthaccept lenCorrect
Qwen3.5+dflash35B-A3B54115417468.04793.622.5344.380335.438TRUE
Qwen3.5+dflash35B-A3B85118542082.121035.0527.2236.73070000700006.562TRUE
Qwen3.5+dflash35B-A3B63116325379.11796.3231.7731.4705405407.667TRUE
Qwen3.5+dflash35B-A3B1391113927488.551569.7127.0237.01020206.524TRUE
Qwen3.5+dflash35B-A3B83118357682.021011.9624.940.15064646TRUE
Qwen3.5+dflash35B-A3B69116932378.12883.2241.8523.89026026010.094TRUE
Qwen3.5+dflash35B-A3B98119836884.961153.4423.542.5601601605.662TRUE
Qwen3.5+dflash35B-A3B1321113260287.281512.4130.0233.31045457.253TRUE
Qwen3.5+dflash35B-A3B88118833482.571065.7327.1136.8904604606.549TRUE
Qwen3.5+dflash35B-A3B88118838882.451067.2639.0625.603663669.463TRUE
Qwen3.5+dflash35B-A3B93119344583.481114.0844.6522.4069469410.854TRUE
Qwen3.5+dflash35B-A3B951195409682.911145.8225.8638.66012136.311FALSE
Qwen3.5+dflash35B-A3B86118661881.811051.2229.7733.59018187.186TRUE
Qwen3.5+dflash35B-A3B80118057179.991000.1132.7730.51060607.931TRUE
Qwen3.5+dflash35B-A3B1221112241287.491394.529.8633.4901251257.228TRUE
Qwen3.5+dflash35B-A3B81118129480.731003.423.8441.9402302305.765TRUE
Qwen3.5+dflash35B-A3B86118644581.741052.0831.7131.54057500575007.672TRUE
Qwen3.5+dflash35B-A3B57115726776.76742.5629.134.360777.026TRUE
Qwen3.5+dflash35B-A3B93119342183.541113.2228.0635.640666.79TRUE
Qwen3.5+dflash35B-A3B91119144482.51102.9932.2431.01015157.789TRUE
Qwen3.5+dflash35B-A3B74117427979.06935.9824.0541.58014145.812TRUE
Qwen3.5+dflash35B-A3B83118322181.321020.6336.5527.360778.84TRUE
Qwen3.5+dflash35B-A3B68116821677.05882.5227.0237.010886.545TRUE
Qwen3.5+dflash35B-A3B67116749975.89882.9130.8232.45026267.448TRUE
Qwen3.5+dflash35B-A3B95119540584.171128.731.5231.730227.642TRUE
Qwen3.5+dflash35B-A3B97119745084.241151.4936.3827.4902432438.824TRUE
Qwen3.5+dflash35B-A3B85118533581.931037.4228.2735.38016166.837TRUE
Qwen3.5+dflash35B-A3B78117826380.03974.6625.9138.6025256.262TRUE
Qwen3.5+dflash35B-A3B97119734383.481161.9727.7436.0501041046.725TRUE
Qwen3.5+dflash35B-A3B67116730277.02869.8535.728.0101091098.629TRUE
Qwen3.5+dflash35B-A3B91119136983.061095.5533.8529.55080808.2TRUE
Qwen3.5+dflash35B-A3B69116918077.47890.6924.8540.23035356TRUE

Qwen3.5_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16))

ModelSizePrefillLen(tokens)Ctx(K)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)ProfilePrefix(tokens)MeasuredOutput(tokens)
Qwen3.5_FQ35b_a3b128161112810082.2491556.24421.02147.572128:1280100
Qwen3.5_FQ35b_a3b250161125010081.7013059.95520.85547.949250:2500100
Qwen3.5_FQ35b_a3b5001611500100131.5043802.17620.91847.805500:5000100
Qwen3.5_FQ35b_a3b100016111000100187.1315343.84921.03847.5331000:10000100
Qwen3.5_FQ35b_a3b200016112000100185.9710754.40220.92347.7952000:20000100
Qwen3.5_FQ35b_a3b250161150010071.7123486.1620.81448.045250:500250100
Qwen3.5_FQ35b_a3b2501611100010070.963523.120.43948.926250:1000750100
Qwen3.5_FQ35b_a3b2501611200010070.8443528.89820.57548.602250:20001750100
Qwen3.5_FQ35b_a3b2501611400010067.6043698.01620.34749.148250:40003750100
Qwen3.5_FQ35b_a3b2501611800010072.4993448.31620.25549.371250:80007750100
Qwen3.5_FQ35b_a3b25016111600010061.2684080.43118.77753.256250:1600015750100

Qwen3.5_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16)) 开启 MTP 投机解码 (GSM8K测试集)

ModelSizePrefillLen(tokens)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)Prompt Indexaccept rate((acclen-1)/(4-1))accept rate(acclen/4)PredictionGround TruthCorrect
Qwen3.5_FQ35b_a3b8691186934092.04944223.1943.1360.76380.82526464TRUE
Qwen3.5_FQ35b_a3b8551185511191.749319.9525.7938.7770.88890.925260260TRUE
Qwen3.5_FQ35b_a3b8841188423792.319576.4923.6542.2780.79050.8464160160TRUE
Qwen3.5_FQ35b_a3b9191191926792.369950.4323.9441.7790.80340.85584545TRUE
Qwen3.5_FQ35b_a3b8751187514792.199490.7624.0941.51100.79840.8547460460TRUE
Qwen3.5_FQ35b_a3b8751187514092.239487.4724.0641.56110.79670.8537366366TRUE
Qwen3.5_FQ35b_a3b8791187918292.39523.1925.5739.11120.87330.91694694TRUE
Qwen3.5_FQ35b_a3b8821188216992.39555.3221.1947.19130.66670.75451213FALSE
Qwen3.5_FQ35b_a3b8721187238792.279450.1624.9440.09140.84710.88761818TRUE
Qwen3.5_FQ35b_a3b8671186718192.249399.1523.642.37150.77780.8386060TRUE
Qwen3.5_FQ35b_a3b9091190915792.459831.8523.4542.65160.7730.8351125125TRUE
Qwen3.5_FQ35b_a3b868118686192.249409.826.8337.27170.91670.9531230230TRUE
Qwen3.5_FQ35b_a3b8721187215192.169461.3225.5839.09180.88620.92075750057500TRUE
Qwen3.5_FQ35b_a3b843118437791.99172.9622.3444.77190.72220.802177TRUE
Qwen3.5_FQ35b_a3b8801188017992.259539.7122.0245.42200.70760.785166TRUE

Qwen3.5_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16)) 开启 DFLASH 投机解码 (GSM8K测试集)

ModelSizePrefillLen(tokens)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)Prefix(tokens)PredictionGround Truthaccept lenCorrect
Qwen3.5_FQ+dflash35B-A3B54115417467.83796.122.3644.720335.438TRUE
Qwen3.5_FQ+dflash35B-A3B85118541582.641028.6228.4935.1070000700006.917TRUE
Qwen3.5_FQ+dflash35B-A3B63116325379.35793.9332.5630.7105405407.906TRUE
Qwen3.5_FQ+dflash35B-A3B1391113927289.431554.3429.3834.03020207.158TRUE
Qwen3.5_FQ+dflash35B-A3B83118346082.441006.8225.9438.55064646.301TRUE
Qwen3.5_FQ+dflash35B-A3B69116932278.65877.2844.1622.65026026010.733TRUE
Qwen3.5_FQ+dflash35B-A3B98119837485.691143.725.6938.9201601606.233TRUE
Qwen3.5_FQ+dflash35B-A3B1321113250088.671488.6728.5335.05045456.944FALSE
Qwen3.5_FQ+dflash35B-A3B88118832983.561053.0926.5937.6104604606.451TRUE
Qwen3.5_FQ+dflash35B-A3B88118843583.581052.9133.1330.1803663668.056TRUE
Qwen3.5_FQ+dflash35B-A3B93119344584.331102.8342.4523.56069469410.349TRUE
Qwen3.5_FQ+dflash35B-A3B95119550084.641122.3522.5944.27013135.495FALSE
Qwen3.5_FQ+dflash35B-A3B86118650082.911037.2624.4840.85018185.952FALSE
Qwen3.5_FQ+dflash35B-A3B80118050081.67979.634.7728.76060608.475TRUE
Qwen3.5_FQ+dflash35B-A3B1221112234989.151368.531.1932.0601251257.587TRUE
Qwen3.5_FQ+dflash35B-A3B81118134381.99987.923.1843.1502302305.623TRUE
Qwen3.5_FQ+dflash35B-A3B86118647283.061035.4128.1135.57057500575006.841TRUE
Qwen3.5_FQ+dflash35B-A3B57115726776.95740.7626.8237.290776.512TRUE
Qwen3.5_FQ+dflash35B-A3B93119350084.491100.7331.0632.20667.576TRUE
Qwen3.5_FQ+dflash35B-A3B91119144483.641087.9832.0231.23015157.789TRUE
Qwen3.5_FQ+dflash35B-A3B74117427979.6929.6126.737.46014146.488TRUE
Qwen3.5_FQ+dflash35B-A3B83118322182.471006.4139.5125.310779.609TRUE
Qwen3.5_FQ+dflash35B-A3B68116821677.98871.9726.1238.290886.353TRUE
Qwen3.5_FQ+dflash35B-A3B67116750077.78861.4130.233.11025257.353TRUE
Qwen3.5_FQ+dflash35B-A3B95119540584.921118.6531.9631.29026267.788TRUE
Qwen3.5_FQ+dflash35B-A3B97119745085.131139.4136.8427.14027279TRUE
Qwen3.5_FQ+dflash35B-A3B85118533482.681028.0128.6534.91028286.958TRUE
Qwen3.5_FQ+dflash35B-A3B78117826381.02962.7327.0736.94029296.575TRUE
Qwen3.5_FQ+dflash35B-A3B97119734084.941141.9727.3936.51030306.667TRUE
Qwen3.5_FQ+dflash35B-A3B67116733877.83860.8835.6528.05031318.667TRUE

Qwen3.6_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16))

ModelSizePrefillLen(tokens)Ctx(K)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)ProfilePrefix(tokens)MeasuredOutput(tokens)
Qwen3.6_FQ35b_a3b128161112810077.5861649.77221.1547.281128:1280100
Qwen3.6_FQ35b_a3b250161125010077.6853218.11120.98447.655250:2500100
Qwen3.6_FQ35b_a3b5001611500100130.3043837.1921.05747.49500:5000100
Qwen3.6_FQ35b_a3b100016111000100199.3385016.61321.0847.4391000:10000100
Qwen3.6_FQ35b_a3b200016112000100200.0059999.73921.20447.1622000:20000100
Qwen3.6_FQ35b_a3b250161150010071.7653483.61121.28646.979250:500250100
Qwen3.6_FQ35b_a3b2501611100010071.7353485.05321.26947.018250:1000750100
Qwen3.6_FQ35b_a3b2501611200010070.4183550.21821.15647.268250:20001750100
Qwen3.6_FQ35b_a3b2501611400010067.1463723.2220.68948.335250:40003750100
Qwen3.6_FQ35b_a3b2501611800010068.4593651.82620.41548.985250:80007750100
Qwen3.6_FQ35b_a3b25016111600010062.793981.53219.57251.093250:1600015750100

Qwen3.6_FQ-35B-A3B (MoE GPTQ,GPTQ INT4 (W4A16)) 开启 DFLASH 投机解码 (GSM8K测试集)

ModelSizePrefillLen(tokens)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)Prefix(tokens)Prompt IndexPredictionGround Truthaccept lenCorrect
Qwen3.6_FQ+dflash35b_a3b54115424953.631006.9831.531.7401336.917TRUE
Qwen3.6_FQ+dflash35b_a3b85118559966.361280.9729.6733.70270000700006.511TRUE
Qwen3.6_FQ+dflash35b_a3b63116323157.911087.9730.9732.29035405406.794TRUE
Qwen3.6_FQ+dflash35b_a3b1391113935377.321797.7630.2733.040420206.66TRUE
Qwen3.6_FQ+dflash35b_a3b83118348165.481267.6226.0838.340564645.726TRUE
Qwen3.6_FQ+dflash35b_a3b69116933860.321143.9346.5721.480626026010.242TRUE
Qwen3.6_FQ+dflash35b_a3b98119850169.881402.4727.536.37071601606.036TRUE
Qwen3.6_FQ+dflash35b_a3b1321113264976.031736.1435.9827.790845457.915TRUE
Qwen3.6_FQ+dflash35b_a3b88118831167.161310.3829.5433.86094604606.479TRUE
Qwen3.6_FQ+dflash35b_a3b88118842079.011113.8145.4222.0201036636610TRUE
Qwen3.6_FQ+dflash35b_a3b93119348780.21159.6447.0721.2401169469410.362TRUE
Qwen3.6_FQ+dflash35b_a3b95119558169.221372.3723.0343.4101212135.052FALSE
Qwen3.6_FQ+dflash35b_a3b86118663066.291297.3836.7727.201318188.077TRUE
Qwen3.6_FQ+dflash35b_a3b80118065676.791041.8539.2725.4601460608.632TRUE
Qwen3.6_FQ+dflash35b_a3b1221112228375.531615.2329.2834.160151251256.432TRUE
Qwen3.6_FQ+dflash35b_a3b81118126664.851249.0532.7230.560162302307.189TRUE
Qwen3.6_FQ+dflash35b_a3b86118641278.531095.1541.6124.0301757500575009.156TRUE
Qwen3.6_FQ+dflash35b_a3b57115727054.981036.8234.1829.26018777.5TRUE
Qwen3.6_FQ+dflash35b_a3b93119354880.271158.5933.2530.07019667.307TRUE
Qwen3.6_FQ+dflash35b_a3b91119143167.991338.5129.7133.6502015156.53TRUE
Qwen3.6_FQ+dflash35b_a3b74117428262.281188.1929.1534.3102114146.409TRUE
Qwen3.6_FQ+dflash35b_a3b83118321577.71068.1539.0925.59022778.6TRUE
Qwen3.6_FQ+dflash35b_a3b68116821959.951134.3443.2623.11023889.522TRUE
Qwen3.6_FQ+dflash35b_a3b67116750059.5112632.131.1502426267.042TRUE
Qwen3.6_FQ+dflash35b_a3b95119538580.851175.0535.0328.54025227.7TRUE
Qwen3.6_FQ+dflash35b_a3b97119757869.711391.5338.6425.880262432438.5TRUE
Qwen3.6_FQ+dflash35b_a3b85118533778.331085.2233.3529.9802716167.326TRUE
Qwen3.6_FQ+dflash35b_a3b78117830663.771223.1734.0329.3902825257.463TRUE
Qwen3.6_FQ+dflash35b_a3b97119732980.861199.5429.9433.40291041046.58TRUE
Qwen3.6_FQ+dflash35b_a3b67116730072.54923.6735.9127.840301091097.895TRUE
Qwen3.6_FQ+dflash35b_a3b91119135379.591143.3635.6628.0403180807.844TRUE
Qwen3.6_FQ+dflash35b_a3b69116920573.25941.9632.1731.0803235357.069TRUE

Qwen3-VL-8B (VLM GPTQ,GPTQ INT4 (W4A16))

ModelSizePrefillLen(tokens)Ctx(K)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)ProfilePrefix(tokens)MeasuredOutput(tokens)
Qwen3VL8b128161112810095.3351342.62913.75272.716128:1280100
Qwen3VL8b2501611250100162.4731538.71713.43774.419250:2500100
Qwen3VL8b5001611500100252.5641979.69411.80784.693500:5000100
Qwen3VL8b100016111000100331.6863014.90411.60286.1881000:10000100
Qwen3VL8b200016112000100314.0926367.56911.4487.4152000:20000100
Qwen3VL8b2501611500100132.331889.2111.69685.497250:500250100
Qwen3VL8b25016111000100134.4111859.9711.39587.76250:1000750100
Qwen3VL8b25016112000100125.4811992.3311.41287.625250:20001750100
Qwen3VL8b25016114000100130.6141914.03511.34288.168250:40003750100
Qwen3VL8b25016118000100105.792363.17410.73393.174250:80007750100
Qwen3VL8b25016111600010076.7643256.7519.754102.522250:1600015750100
ModelSizePrefillLen(k)Ctx(k)BatchNChipInput(k)Output(k)Prefill(tps)Decode(tps)Vision(fps)Precision
Qwen3-VL8B0.2516110.250.05161.49913.4636.004w4a16
Qwen3-VL8B0.2516110.50.05137.50310.396.004w4a16
Qwen3-VL8B0.25161110.05105.34710.8566.004w4a16
Qwen3-VL8B0.25161120.05124.8189.756.004w4a16
Qwen3-VL8B0.25161140.05130.5479.746.004w4a16
Qwen3-VL8B0.25161180.05105.7199.7466.004w4a16
Qwen3-VL8B0.251611160.0576.7289.7526.004w4a16

Qwen3.5 additional table

ModelSizePrefillLen(tokens)Ctx(K)BatchNChipInput(tokens)Output(tokens)Prefill(tps)TTFT(ms)Decode(tps)TPOT(ms)ProfilePrefix(tokens)
Qwen3.535b_a3b128161112810050.12554.9115.24965.579128:1280
Qwen3.535b_a3b250161125010079.993125.37315.27865.454250:2500
Qwen3.535b_a3b5001611500100131.0153816.35114.97866.765500:5000
Qwen3.535b_a3b100016111000100208.7984789.32314.81767.4921000:10000
Qwen3.535b_a3b200016112000100219.029131.57114.76267.7412000:20000
Qwen3.535b_a3b250161125010080.1153120.50814.96966.803250:500250
Qwen3.535b_a3b2501611100010079.9033128.79314.81467.503250:1000750
Qwen3.535b_a3b2501611200010078.8083172.25714.77867.67250:20001750
Qwen3.535b_a3b2501611400010069.8683578.15214.76867.713250:40003750
Qwen3.535b_a3b2501611800010068.9633625.12714.70568.002250:80007750
Qwen3.535b_a3b25016111600010061.8074044.86514.3869.541250:1600015750