# vLLM runtime patch (ships with this MXFP4 checkpoint)

## Why
This checkpoint quantizes the routed experts to **MXFP4** (the rest stays native
MXFP8). MiniMax-M3 uses the clamped **SwiGLU-OAI** activation
(`swiglu_limit: 7.0` in `text_config`).

The stock `vllm/vllm-openai:minimax-m3` image has a bug in the compressed-tensors
MXFP4 MoE method: `get_fused_moe_quant_config()` builds the kernel quant config
**without forwarding the SwiGLU clamp params**, so `gemm1_clamp_limit` stays
`None` and loading fails with:

```
AssertionError: SWIGLUOAI_UNINTERLEAVE requires clamp_limit
```

This is purely a runtime bug — the checkpoint and its `config.json` are correct
(`swiglu_limit=7.0` is reachable via `hf_text_config`).

## The fix
`compressed_tensors_moe_w4a4_mxfp4.py` here is the patched file. It forwards
`swiglu_limit / swiglu_alpha / swiglu_beta` (read from the already-populated MoE
config) into the kernel quant config:

- **W4A16 (weight-only, this checkpoint's path):** pass `gemm1_alpha`,
  `gemm1_beta`, `swiglu_limit` to `make_mxfp4_moe_quant_config` — which maps
  `gemm1_clamp_limit = swiglu_limit` per backend.
- **W4A4 (cutlass):** set the fields on the returned config directly (that
  builder doesn't expose them).

Send the same change upstream to the fork.

## How to apply (no image rebuild)
Bind-mount the file over the buggy one. Path is for the image's Python 3.12:

```bash
docker run --gpus all --privileged --ipc=host -p 8000:8000 \
  -v /mnt/storage:/root/.cache/huggingface \
  -v /mnt/storage/M3-MXFP4/vllm_patch/compressed_tensors_moe_w4a4_mxfp4.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py \
  vllm/vllm-openai:minimax-m3 /root/.cache/huggingface/M3-MXFP4 \
  --block-size 128 \
  --tool-call-parser minimax_m3 --enable-auto-tool-choice \
  --reasoning-parser minimax_m3 \
  --load-format fastsafetensors
```

Verify after startup: the model loads past the activation assertion.
If you bump the VRAM ceiling, add `--gpu-memory-utilization 0.85` and a smaller
`--max-model-len` (the 1M default makes the KV cache huge).

Patched against image `vllm/vllm-openai:minimax-m3` (vllm 0.1.dev17492+g454b47db8).