From 6f70a3d270e18d4f6610ab8749931a2b8e5de430 Mon Sep 17 00:00:00 2001 From: Faradawn Yang <73060648+faradawn@users.noreply.github.com> Date: Mon, 3 Aug 2026 06:57:37 -0700 Subject: [PATCH 1/3] Update MiniMax M3 for B300 NVFP4 Signed-off-by: Faradawn Yang <73060648+faradawn@users.noreply.github.com> --- models/MiniMaxAI/MiniMax-M3.yaml | 30 +++++++++++++++++++++++++++++- 1 file changed, 29 insertions(+), 1 deletion(-) diff --git a/models/MiniMaxAI/MiniMax-M3.yaml b/models/MiniMaxAI/MiniMax-M3.yaml index 8bbccf883..6318117ed 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: 2026-08-03 difficulty: advanced tasks: - text @@ -14,6 +14,7 @@ meta: hardware: h200: verified b200: verified + b300: verified mi300x: verified mi355x: verified @@ -161,6 +162,11 @@ variants: description: "NVIDIA-quantized NVFP4 weights (ModelOpt) for Blackwell (B200/B300) — native FP4 tensor cores, ~half the VRAM of MXFP8." extra_args: - "--trust-remote-code" + # FlashInfer TRT-LLM attention kernels with an FP8 indexer KV cache. + - "--attention_config.backend" + - "FLASHINFER" + - "--attention_config.use_trtllm_attention" + - "true" - "--attention_config.indexer_kv_dtype" - "fp8" extra_env: @@ -649,6 +655,28 @@ guide: | --speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3-GQA", "num_speculative_tokens": 3, "attention_backend": "FLASH_ATTN"}' ``` + ### Agentic coding on B300 + + The validated B300 agentic-coding lane runs the GQA draft head at TP4 with + thinking always on, and pins thinking server-side so every turn — including + turns after tool results — reasons before answering: + + ```bash + VLLM_FLOAT32_MATMUL_PRECISION=high \ + vllm serve nvidia/MiniMax-M3-NVFP4 \ + --tensor-parallel-size 4 \ + --trust-remote-code \ + --block-size 128 \ + --language-model-only \ + --enable-prefix-caching \ + --attention-config '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' \ + --reasoning-parser minimax_m3 \ + --tool-call-parser minimax_m3 \ + --enable-auto-tool-choice \ + --default-chat-template-kwargs '{"thinking_mode": "enabled"}' \ + --speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3-GQA", "num_speculative_tokens": 3, "attention_backend": "FLASH_ATTN"}' + ``` + ## Quantized Variant (MXFP4 on MI355X) [`amd/MiniMax-M3-MXFP4`](https://huggingface.co/amd/MiniMax-M3-MXFP4) From acbb453c0bc6b6180b7c16e284ff6b1582b3d3fb Mon Sep 17 00:00:00 2001 From: Faradawn Yang <73060648+faradawn@users.noreply.github.com> Date: Wed, 5 Aug 2026 09:11:55 -0700 Subject: [PATCH 2/3] Update MiniMax M3 for B300 NVFP4 Signed-off-by: Faradawn Yang <73060648+faradawn@users.noreply.github.com> --- models/MiniMaxAI/MiniMax-M3.yaml | 29 +++++++---------------------- 1 file changed, 7 insertions(+), 22 deletions(-) diff --git a/models/MiniMaxAI/MiniMax-M3.yaml b/models/MiniMaxAI/MiniMax-M3.yaml index 6318117ed..45d80f39a 100644 --- a/models/MiniMaxAI/MiniMax-M3.yaml +++ b/models/MiniMaxAI/MiniMax-M3.yaml @@ -73,6 +73,11 @@ features: description: "Skip loading the vision encoder for text-only workloads — frees VRAM for KV cache. Mutually exclusive with encoder_parallel." args: - "--language-model-only" + thinking_always_on: + description: "Pin `thinking_mode: enabled` server-side — every turn reasons before answering, including after tool results. Recommended for agentic workflows." + args: + - "--default-chat-template-kwargs" + - '{"thinking_mode": "enabled"}' encoder_parallel: description: "Run the vision encoder data-parallel instead of tensor-parallel — avoids TP comm overhead on the small encoder. Also enables the host-shared-memory multimodal processor cache and the per-brand encoder attention backend (FlashInfer on NVIDIA, AITER FlashAttention on AMD). Mutually exclusive with text_only." # No platform-agnostic `args`: this recipe targets only NVIDIA + AMD, both @@ -101,6 +106,7 @@ features: opt_in_features: - spec_decoding - text_only + - thinking_always_on - encoder_parallel variants: @@ -171,6 +177,7 @@ variants: - "fp8" extra_env: VLLM_FLOAT32_MATMUL_PRECISION: "high" + VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm" mxfp4: model_id: "amd/MiniMax-M3-MXFP4" precision: mxfp4 @@ -655,28 +662,6 @@ guide: | --speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3-GQA", "num_speculative_tokens": 3, "attention_backend": "FLASH_ATTN"}' ``` - ### Agentic coding on B300 - - The validated B300 agentic-coding lane runs the GQA draft head at TP4 with - thinking always on, and pins thinking server-side so every turn — including - turns after tool results — reasons before answering: - - ```bash - VLLM_FLOAT32_MATMUL_PRECISION=high \ - vllm serve nvidia/MiniMax-M3-NVFP4 \ - --tensor-parallel-size 4 \ - --trust-remote-code \ - --block-size 128 \ - --language-model-only \ - --enable-prefix-caching \ - --attention-config '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' \ - --reasoning-parser minimax_m3 \ - --tool-call-parser minimax_m3 \ - --enable-auto-tool-choice \ - --default-chat-template-kwargs '{"thinking_mode": "enabled"}' \ - --speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3-GQA", "num_speculative_tokens": 3, "attention_backend": "FLASH_ATTN"}' - ``` - ## Quantized Variant (MXFP4 on MI355X) [`amd/MiniMax-M3-MXFP4`](https://huggingface.co/amd/MiniMax-M3-MXFP4) From f072dab3b1c7f8583191467ca98b37627995afeb Mon Sep 17 00:00:00 2001 From: Faradawn Yang <73060648+faradawn@users.noreply.github.com> Date: Wed, 5 Aug 2026 09:23:56 -0700 Subject: [PATCH 3/3] Update MiniMax M3 for B300 NVFP4 Signed-off-by: Faradawn Yang <73060648+faradawn@users.noreply.github.com> --- models/MiniMaxAI/MiniMax-M3.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/models/MiniMaxAI/MiniMax-M3.yaml b/models/MiniMaxAI/MiniMax-M3.yaml index 45d80f39a..31804dcf4 100644 --- a/models/MiniMaxAI/MiniMax-M3.yaml +++ b/models/MiniMaxAI/MiniMax-M3.yaml @@ -168,6 +168,8 @@ variants: description: "NVIDIA-quantized NVFP4 weights (ModelOpt) for Blackwell (B200/B300) — native FP4 tensor cores, ~half the VRAM of MXFP8." extra_args: - "--trust-remote-code" + - "--kv-cache-dtype" + - "fp8" # FlashInfer TRT-LLM attention kernels with an FP8 indexer KV cache. - "--attention_config.backend" - "FLASHINFER"