From 60548501bb13fcceeee1d085356b7cad02874f97 Mon Sep 17 00:00:00 2001 From: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com> Date: Tue, 1 Sep 2026 14:14:30 +0800 Subject: [PATCH] [Docs] Add NVFP4 section to GLM-5.3-Flash cookbook (#37109) --- .../autoregressive/GLM/GLM-5.3-Flash.mdx | 1 - .../zai-org/glm-5.3-flash-benchmarks.jsx | 84 +++++- .../configs/zai-org/glm-5.3-flash.jsx | 259 +++++++++++++++++- 3 files changed, 324 insertions(+), 20 deletions(-) diff --git a/docs/cookbook/autoregressive/GLM/GLM-5.3-Flash.mdx b/docs/cookbook/autoregressive/GLM/GLM-5.3-Flash.mdx index 9124e1b5c..b6c918c7f 100644 --- a/docs/cookbook/autoregressive/GLM/GLM-5.3-Flash.mdx +++ b/docs/cookbook/autoregressive/GLM/GLM-5.3-Flash.mdx @@ -184,7 +184,6 @@ sglang serve \ --max-running-requests 64 \ --chunked-prefill-size 8192 \ --max-prefill-tokens 8192 \ - --disable-shared-experts-fusion \ --disable-prefill-cuda-graph \ --speculative-algorithm EAGLE \ --speculative-num-steps 3 \ diff --git a/docs/src/snippets/configs/zai-org/glm-5.3-flash-benchmarks.jsx b/docs/src/snippets/configs/zai-org/glm-5.3-flash-benchmarks.jsx index e433138b2..b87bac937 100644 --- a/docs/src/snippets/configs/zai-org/glm-5.3-flash-benchmarks.jsx +++ b/docs/src/snippets/configs/zai-org/glm-5.3-flash-benchmarks.jsx @@ -22,7 +22,7 @@ export const benchmarks = [ "Measured on 4x GB300 (TP4/EP4) with the final weights (zai-org/GLM-5.3-Flash, c5b82b63e37b) on the current release-image tree (d6ab04bdf1), adaptive MTP 5/1/6 with SGLANG_SIMULATE_ACC_LEN=3 (accept length confirmed 3.00 in the bench summary and server log): 80 random requests at 1,024 input / 256 output tokens and concurrency 16 produced 1,824.66 aggregate output tok/s after two discarded warmups. Simulated accept length makes this a throughput-mechanism number. Accuracy is from the shared non-simulated full GSM8K gate: 97.50% with a 100% stop rate over all 1,319 problems.", }, { - match: { hw: "gb300", strategy: "low-latency", kvDsaPair: "fp8-trtllm" }, + match: { hw: "gb300", strategy: "low-latency", kvDsaPair: "fp8-trtllm", quant: "fp8" }, sglang_version: "d6ab04bdf1", latencyPercentile: "Mean", speed: [ @@ -43,7 +43,7 @@ export const benchmarks = [ "The Low Latency recipe with FP8 KV + TRT-LLM DSA on 4x GB300, final weights (c5b82b63e37b) on the current release-image tree (d6ab04bdf1), adaptive MTP 5/1/6 with SGLANG_SIMULATE_ACC_LEN=3 (accept 3.00): 80 random requests at 1,024 input / 256 output tokens and concurrency 16 produced 1,853.88 aggregate output tok/s — 1.6% above the BF16 + TileLang Low Latency row, with mean TPOT 6.25 ms vs 6.43 ms. Draft and target full-graph capture succeeded for this combination. The speed rows were measured with the NEXTN spelling and --disable-shared-experts-fusion, which resolve to the same runtime path as the published command on this tree.", }, { - match: { hw: "gb300", strategy: "low-latency", kvDsaPair: "fp8-trtllm", dcp: "4" }, + match: { hw: "gb300", strategy: "low-latency", kvDsaPair: "fp8-trtllm", dcp: "4", quant: "fp8" }, sglang_version: "d6ab04bdf1", latencyPercentile: "Mean", speed: [ @@ -64,7 +64,7 @@ export const benchmarks = [ "The Low Latency recipe with FP8 KV + TRT-LLM DSA and DCP4 (--dcp-size 4 --dcp-comm-backend a2a --dcp-replicate-q-proj) on 4x GB300, final weights (c5b82b63e37b) on the d6ab04bdf1 tree, adaptive MTP 5/1/6 with full decode graph: 80 random requests at 1,024 input / 256 output tokens and concurrency 16 produced 1,680.61 aggregate output tok/s at a 3.937 accept length — about 10% below the non-DCP FP8 Low Latency row. TRT-LLM DSA DCP decode returns the LSE natively, so this arm needs no patch.", }, { - match: { hw: "gb300", strategy: "low-latency", kvDsaPair: "bf16-tilelang", dcp: "4" }, + match: { hw: "gb300", strategy: "low-latency", kvDsaPair: "bf16-tilelang", dcp: "4", quant: "fp8" }, sglang_version: "d6ab04bdf1", latencyPercentile: "Mean", speed: [ @@ -131,7 +131,7 @@ export const benchmarks = [ "Measured on 4x GB300 (TP4/EP4) with the final weights (zai-org/GLM-5.3-Flash, c5b82b63e37b) on the current release-image tree (d6ab04bdf1), speculative decoding off, after two discarded warmups per row: 1,161.22 / 2,660.24 / 4,828.33 aggregate output tok/s at concurrency 16 / 64 / 256 (80 / 320 / 1,280 random requests at 1,024 input / 256 output tokens). The server ran exactly the published cell command. Throughput at 256 is still scaling but sublinear (prefill queueing). Accuracy is from the shared non-simulated full GSM8K gate: 97.50% with a 100% stop rate over all 1,319 problems. With HiCache L1+L2 (32 GB host tier, 16k prefill chunks) the same protocol measured 1,202.07 / 2,696.20 / 4,634.47 tok/s — within 4% of the non-HiCache rows; the random dataset has no prefix reuse, so L2 benefit was not exercised.", }, { - match: { hw: "gb300", strategy: "high-throughput", kvDsaPair: "fp8-trtllm" }, + match: { hw: "gb300", strategy: "high-throughput", kvDsaPair: "fp8-trtllm", quant: "fp8" }, sglang_version: "d6ab04bdf1", latencyPercentile: "Mean", speed: [ @@ -176,6 +176,76 @@ export const benchmarks = [ notes: "FP8 KV cache with TRT-LLM DSA on 4x GB300, final weights (c5b82b63e37b) on the current release-image tree (d6ab04bdf1), same protocol as the BF16 rows: 1,227.07 / 2,738.61 / 4,977.02 aggregate output tok/s at concurrency 16 / 64 / 256 — 2.9–5.7% above BF16 + TileLang across the curve, and the FP8 pool holds 12.6M tokens per rank vs 7.0M at BF16 (1.8x capacity at identical pool bytes). Accuracy is the full GSM8K gate on this variant: 97.35% vs 97.50% on BF16 KV, a 0.15-point gap inside sampling noise, with a 100% stop rate over all 1,319 problems. With HiCache L1+L2 (32 GB host tier, 16k prefill chunks) the same protocol measured 1,263.85 / 2,763.31 / 4,773.95 tok/s — within 5% of the non-HiCache rows; the random dataset has no prefix reuse, so L2 benefit was not exercised.", }, + { + match: { hw: "gb300", strategy: "low-latency", quant: "nvfp4", kvDsaPair: "bf16-tilelang" }, + sglang_version: "033446bb05", + latencyPercentile: "Mean", + speed: [ + { + workload: { + dataset: "random", + isl: 1024, + osl: 256, + max_concurrency: 16, + num_prompts: 80, + }, + ttft_ms: 242.66, + tpot_ms: 16.95, + tokens_per_sec_per_gpu: 1022.61, + }, + ], + accuracy: { gsm8k_pct: 97.14, aime2026_pct: 92.45 }, + notes: + "RadixArk/GLM-5.3-Flash-NVFP4 — NVFP4 W4A4 post-training quantization of zai-org/GLM-5.3-Flash-BF16 with NVIDIA Model Optimizer 0.46.0 (abs-max scaling, group size 16): routed and shared experts plus the dense MLPs are FP4, while all attention (KDA, DSA indexer, MLA), the router, norms, the vision tower, the MTP layer, embeddings, and the LM head stay BF16. Measured on 4x GB300 with the lmsysorg/sglang:glm-5.3-flash image (PR #36507 head 033446bb05), adaptive MTP 5/1/6, BF16 KV + TileLang DSA, and the flashinfer_cutlass MoE runner. GSM8K 97.14% over the full 1,319-example split x 4 seeds (per-seed range 96.89-97.42%, stop rate 99.85-100%) and AIME 2026 92.45% (30 problems x 16 repeats x 4 seeds = 1,920 generations, per-seed range 91.67-93.54%), both at temperature 1.0 / top_p 0.95. The accuracy runs used the NEXTN spelling of --speculative-algorithm, which resolves to the same runtime path as the published EAGLE command on this tree. The speed row was measured on the unpatched docker tree with SGLANG_SIMULATE_ACC_LEN=3 pinning the accept length (confirmed 2.98): 80 random requests at 1,024 input / 256 output tokens and concurrency 16 produced 818.09 aggregate output tok/s after two discarded warmups. Simulated accept length makes this a throughput-mechanism number; a same-pod A/B against the SwiGLU-fusion-disabled tree landed within ±3%, so the unpatched numbers are the published ones. The companion GSM8K gate on this stack scored 97.27% with a 99.92% stop rate over all 1,319 problems. The gap to the fp8 rows (~45%) is recipe-level, not kernel-isolated: flashinfer_cutlass + BF16 KV + TileLang DSA + TP4 without EP vs the fp8 cells' deep_gemm + FP8 KV + TRT-LLM DSA + EP4.", + }, + { + match: { hw: "gb300", strategy: "high-throughput", quant: "nvfp4", kvDsaPair: "bf16-tilelang" }, + sglang_version: "033446bb05", + latencyPercentile: "Mean", + speed: [ + { + workload: { + dataset: "random", + isl: 1024, + osl: 256, + max_concurrency: 16, + num_prompts: 80, + }, + ttft_ms: 188.6, + tpot_ms: 22.01, + tokens_per_sec_per_gpu: 786.79, + }, + { + workload: { + dataset: "random", + isl: 1024, + osl: 256, + max_concurrency: 64, + num_prompts: 320, + }, + ttft_ms: 255.67, + tpot_ms: 56.88, + tokens_per_sec_per_gpu: 1317.68, + }, + { + workload: { + dataset: "random", + isl: 1024, + osl: 256, + max_concurrency: 256, + num_prompts: 1280, + }, + ttft_ms: 413.08, + tpot_ms: 123.03, + tokens_per_sec_per_gpu: 2518.95, + }, + ], + accuracy: { gsm8k_pct: 97.14, aime2026_pct: 92.45 }, + notes: + "RadixArk/GLM-5.3-Flash-NVFP4 with speculative decoding off — same checkpoint, image, and 4x GB300 measurement stack as the NVFP4 Low Latency row (ModelOpt 0.46.0 NVFP4 W4A4, abs-max, group size 16; MoE and dense MLPs in FP4, attention/router/MTP/embeddings BF16). Accuracy is a checkpoint-level result carried from that arm: GSM8K 97.14% over the full 1,319-example split x 4 seeds (per-seed range 96.89-97.42%, stop rate 99.85-100%) and AIME 2026 92.45% (30 problems x 16 repeats x 4 seeds, per-seed range 91.67-93.54%). Those runs used the NEXTN spelling of --speculative-algorithm on the adaptive-MTP arm, which resolves to the same runtime path as the published EAGLE command. The speed rows were measured on the unpatched docker tree after two discarded warmups per row: 629.43 / 1,054.14 / 2,015.16 aggregate output tok/s at concurrency 16 / 64 / 256 (80 / 320 / 1,280 random requests at 1,024 input / 256 output tokens), with decode on CUDA graphs through bs256 — the published cell command carries no --cuda-graph-max-bs cap. The companion GSM8K gate on this stack scored 97.27% with a 99.92% stop rate over all 1,319 problems. The gap to the fp8 rows (~40-55%) is recipe-level, not kernel-isolated: flashinfer_cutlass + BF16 KV + TileLang DSA + TP4 without EP vs the fp8 cells' deep_gemm + FP8 KV + TRT-LLM DSA + EP4.", + }, + { match: { hw: "gb300", strategy: "low-latency", quant: "nvfp4", kvDsaPair: "fp8-trtllm" } }, + { match: { hw: "gb300", strategy: "high-throughput", quant: "nvfp4", kvDsaPair: "fp8-trtllm" } }, { match: { hw: "h100", strategy: "low-latency" }, sglang_version: "f040cc72e6", @@ -234,6 +304,12 @@ export const benchmarks = [ }, { match: { hw: "gb200", strategy: "low-latency" } }, { match: { hw: "gb200", strategy: "high-throughput" } }, + { match: { hw: "gb200", strategy: "low-latency", quant: "nvfp4" } }, + { match: { hw: "gb200", strategy: "high-throughput", quant: "nvfp4" } }, + { match: { hw: "b200", strategy: "low-latency", quant: "nvfp4" } }, + { match: { hw: "b200", strategy: "high-throughput", quant: "nvfp4" } }, + { match: { hw: "b300", strategy: "low-latency", quant: "nvfp4" } }, + { match: { hw: "b300", strategy: "high-throughput", quant: "nvfp4" } }, { match: { hw: "mi300x", strategy: "high-throughput" }, sglang_version: "9e692c9216", diff --git a/docs/src/snippets/configs/zai-org/glm-5.3-flash.jsx b/docs/src/snippets/configs/zai-org/glm-5.3-flash.jsx index e968a07fc..407d56599 100644 --- a/docs/src/snippets/configs/zai-org/glm-5.3-flash.jsx +++ b/docs/src/snippets/configs/zai-org/glm-5.3-flash.jsx @@ -21,6 +21,19 @@ export const config = { { id: "high-throughput", label: "High Throughput", subtitle: "Spec decode off" }, ], }, + { + id: "quant", + title: "Quantization", + options: [ + { id: "fp8", label: "FP8" }, + { + id: "nvfp4", + label: "NVFP4", + disabled: (s) => !["gb300", "gb200", "b200", "b300"].includes(s.hw), + disableReason: "The NVFP4 W4A4 kernels are Blackwell-only; Hopper and AMD ROCm cannot serve this checkpoint.", + }, + ], + }, ], isRecommendedSelection(s) { @@ -123,6 +136,7 @@ export const config = { modelNames: { default: "zai-org/GLM-5.3-Flash", + nvfp4: "RadixArk/GLM-5.3-Flash-NVFP4", }, placeholders: { @@ -169,6 +183,7 @@ sgl-eval run gsm8k \\ accuracyLabels: [ ["gsm8k_pct", "GSM8K", "%"], + ["aime2026_pct", "AIME 2026", "%"], ], // Support is not in a public sglang release yet. NVIDIA uses the @@ -337,7 +352,7 @@ sgl-eval run gsm8k \\ cells: [ { - match: { hw: "gb300", strategy: "low-latency" }, + match: { hw: "gb300", strategy: "low-latency", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -368,7 +383,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "gb300", strategy: "high-throughput" }, + match: { hw: "gb300", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -393,8 +408,222 @@ sgl-eval run gsm8k \\ "--port {{PORT}}", ], }, + // RadixArk NVFP4 W4A4 checkpoint (ModelOpt 0.46.0, abs-max, group size + // 16): routed and shared experts plus the dense MLPs are FP4; attention, + // router, MTP, embeddings, and the vision tower stay BF16. Validated on + // 4x GB300 on the stock image with both KV/DSA pairings: the speed rows + // were measured with BF16 KV + TileLang DSA, while FP8 KV + TRT-LLM DSA + // passed smoke, a 200-example GSM8K check, a 600-request soak, and the + // TB2.1 run without separate speed measurements. { - match: { hw: "h100", strategy: "low-latency" }, + match: { hw: "gb300", strategy: "low-latency", quant: "nvfp4" }, + nnodes: 1, + verified: true, + verificationStatus: (s) => + ["bf16-tilelang", "fp8-trtllm"].includes(s.kvDsaPair) && + s.mmTransport === "auto" && + s.hicache === "off" && + s.dcp === "off" + ? "verified" + : "unverified", + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 4", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--speculative-algorithm EAGLE", + "--speculative-num-steps 5", + "--speculative-eagle-topk 1", + "--speculative-num-draft-tokens 6", + "--speculative-adaptive", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--cuda-graph-max-bs 32", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "gb300", strategy: "high-throughput", quant: "nvfp4" }, + nnodes: 1, + verified: true, + verificationStatus: (s) => + ["bf16-tilelang", "fp8-trtllm"].includes(s.kvDsaPair) && + s.mmTransport === "auto" && + s.hicache === "off" && + s.dcp === "off" + ? "verified" + : "unverified", + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 4", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + // Same NVFP4 recipe on the remaining Blackwell platforms, at each fp8 + // cell's parallelism (gb200 TP4/EP4, b200/b300 TP8/EP8). Not measured on + // this hardware, so every cell here reports unverified. + { + match: { hw: "gb200", strategy: "low-latency", quant: "nvfp4" }, + nnodes: 1, + verified: false, + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 4", + "--ep-size 4", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--speculative-algorithm EAGLE", + "--speculative-num-steps 5", + "--speculative-eagle-topk 1", + "--speculative-num-draft-tokens 6", + "--speculative-adaptive", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--cuda-graph-max-bs 32", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "gb200", strategy: "high-throughput", quant: "nvfp4" }, + nnodes: 1, + verified: false, + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 4", + "--ep-size 4", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "b200", strategy: "low-latency", quant: "nvfp4" }, + nnodes: 1, + verified: false, + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 8", + "--ep-size 8", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--speculative-algorithm EAGLE", + "--speculative-num-steps 5", + "--speculative-eagle-topk 1", + "--speculative-num-draft-tokens 6", + "--speculative-adaptive", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--cuda-graph-max-bs 32", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "b200", strategy: "high-throughput", quant: "nvfp4" }, + nnodes: 1, + verified: false, + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 8", + "--ep-size 8", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "b300", strategy: "low-latency", quant: "nvfp4" }, + nnodes: 1, + verified: false, + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 8", + "--ep-size 8", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--speculative-algorithm EAGLE", + "--speculative-num-steps 5", + "--speculative-eagle-topk 1", + "--speculative-num-draft-tokens 6", + "--speculative-adaptive", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--cuda-graph-max-bs 32", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "b300", strategy: "high-throughput", quant: "nvfp4" }, + nnodes: 1, + verified: false, + env: [], + flags: [ + "--model-path {{MODEL_NAME}}", + "--quantization modelopt_fp4", + "--tp-size 8", + "--ep-size 8", + "--dsa-prefill-backend tilelang", + "--dsa-decode-backend tilelang", + "--kv-cache-dtype bfloat16", + "--moe-runner-backend flashinfer_cutlass", + "--reasoning-parser glm45", + "--tool-call-parser glm47", + "--mem-fraction-static 0.85", + "--host {{HOST_IP}}", + "--port {{PORT}}", + ], + }, + { + match: { hw: "h100", strategy: "low-latency", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -423,7 +652,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "h100", strategy: "high-throughput" }, + match: { hw: "h100", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -445,7 +674,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "h200", strategy: "low-latency" }, + match: { hw: "h200", strategy: "low-latency", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -474,7 +703,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "h200", strategy: "high-throughput" }, + match: { hw: "h200", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -495,7 +724,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "b200", strategy: "low-latency" }, + match: { hw: "b200", strategy: "low-latency", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => (s.hicache === "off" ? "verified" : "unverified"), @@ -520,7 +749,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "b200", strategy: "high-throughput" }, + match: { hw: "b200", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -541,7 +770,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "b300", strategy: "low-latency" }, + match: { hw: "b300", strategy: "low-latency", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => (s.hicache === "off" ? "verified" : "unverified"), @@ -566,7 +795,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "b300", strategy: "high-throughput" }, + match: { hw: "b300", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, verificationStatus: (s) => @@ -587,7 +816,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "gb200", strategy: "low-latency" }, + match: { hw: "gb200", strategy: "low-latency", quant: "fp8" }, nnodes: 1, verified: false, env: [], @@ -611,7 +840,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "gb200", strategy: "high-throughput" }, + match: { hw: "gb200", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: false, env: [], @@ -635,7 +864,7 @@ sgl-eval run gsm8k \\ // while Triton owns the MoE runner. CUDA graphs stay disabled because that // is the architecture-gated configuration used for correctness validation. { - match: { hw: "mi300x", strategy: "high-throughput" }, + match: { hw: "mi300x", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, env: ["SGLANG_USE_AITER=1"], @@ -655,7 +884,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "mi325x", strategy: "high-throughput" }, + match: { hw: "mi325x", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: false, warn: "This MI325X recipe is inferred from the validated MI300X gfx942 path. It has not been measured directly on MI325X.", @@ -676,7 +905,7 @@ sgl-eval run gsm8k \\ ], }, { - match: { hw: "mi355x", strategy: "high-throughput" }, + match: { hw: "mi355x", strategy: "high-throughput", quant: "fp8" }, nnodes: 1, verified: true, env: ["SGLANG_USE_AITER=1"],