From 0866a8242f6771eccf60043359a7ddfd2c78d2dc Mon Sep 17 00:00:00 2001 From: Yash Sawant Date: Tue, 4 Aug 2026 23:09:26 -0700 Subject: [PATCH] fix(MiniMax-M3): gate indexer_kv_dtype fp8 to Blackwell for MXFP8 Signed-off-by: Yash Sawant --- models/MiniMaxAI/MiniMax-M3.yaml | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/models/MiniMaxAI/MiniMax-M3.yaml b/models/MiniMaxAI/MiniMax-M3.yaml index c60a7f20..cf3e3d9f 100644 --- a/models/MiniMaxAI/MiniMax-M3.yaml +++ b/models/MiniMaxAI/MiniMax-M3.yaml @@ -4,7 +4,7 @@ meta: provider: "MiniMax" description: "MiniMax M3 vision-language MoE (427B total / 26B active) for frontier coding, agent toolchains, and 1M-token reasoning via MSA sparse attention — native multimodal (image + video + computer use); BF16 plus MXFP8, NVIDIA Blackwell NVFP4, and AMD MI355X MXFP4 variants. Runs on NVIDIA (Hopper/Blackwell) and AMD CDNA4/CDNA3." date_added: 2026-06-12 - date_updated: 2026-07-15 + date_updated: 2026-08-04 difficulty: advanced tasks: - text @@ -118,11 +118,12 @@ variants: # comfortably, ~4 GPUs for weights alone on Blackwell (B200/B300) or AMD MI350X/MI355X (gfx950)). vram_minimum_gb: 513 description: "NVIDIA-quantized MXFP8 weights — Blackwell (B200/B300) for native MX tensor cores, and AMD CDNA4 (MI350X/MI355X, gfx950) for native MXFP8 Matrix Cores." - extra_args: - - "--attention_config.indexer_kv_dtype" - - "fp8" hardware_overrides: blackwell: + # FP8 index cache needs the SM100 MSA indexer; Triton (Hopper/AMD) is bf16-only. + extra_args: + - "--attention_config.indexer_kv_dtype" + - "fp8" extra_env: VLLM_FLOAT32_MATMUL_PRECISION: "high" h100: @@ -745,6 +746,9 @@ guide: | - **`--block-size` mismatch.** MSA's sparse block size is 128; the vLLM KV cache block size must match. Using the default (16) misaligns the sparse attention indexing (on AMD it crashes with `No common block size for 16`). + - **`indexer_kv_dtype=fp8` on Hopper/AMD.** The Triton indexer only supports + bf16. Drop `--attention_config.indexer_kv_dtype fp8` on non-Blackwell GPUs + (the command builder already gates it to Blackwell for the mxfp8 variant). - **Parsers.** `--tool-call-parser` and `--reasoning-parser` both use `minimax_m3` — distinct from `minimax_m2` used by earlier releases. - **Long context KV cache.** See *Context Length & GPU Memory* above — cap