Property	Value
Format	OCP MX FP8 (E4M3)
Block size	32 elements
Scale dtype	uint8 (E8M0)
Weight quantization	Static, per 32-element group
Activation quantization	Dynamic, per 32-element group
Layers quantized	All `nn.Linear` in the 80 decoder layers (560 total)
Layers skipped	`lm_head`, `embed_tokens`
Checkpoint format	`compressed-tensors` (auto-detected by vLLM)

The quantized layers cover all attention projections (q_proj, k_proj, v_proj, o_proj) and MLP projections (gate_proj, up_proj, down_proj) in every decoder layer.

Hardware requirements

MXFP8 hardware acceleration requires AMD MI350X / MI355X (CDNA4, gfx950) with ROCm ≥ 7.2.0, or NVIDIA Blackwell (SM100+).

Note: This checkpoint will NOT run correctly on AMD MI300X (gfx942) or NVIDIA Hopper (H100/H200). Those GPUs use different FP8 encodings (FNUZ and E4M3FNUZ respectively) that are incompatible with the OCP MX format stored here.

Usage with vLLM

python

from vllm import LLM, SamplingParams

llm = LLM(
    model="talumbau/Qwen2.5-72B-Instruct-mxfp8",
    # vLLM auto-detects compressed-tensors format — no --quantization flag needed
)

outputs = llm.generate(
    ["Tell me about AMD MI350X"],
    SamplingParams(max_tokens=200, temperature=0.7),
)
print(outputs[0].outputs[0].text)

Or from the command line:

vllm serve talumbau/Qwen2.5-72B-Instruct-mxfp8

Comparison variants

This checkpoint is part of a three-format study:

Variant	Repo	Format
bf16 (original)	`Qwen/Qwen2.5-72B-Instruct`	bfloat16
fp8 (per-channel/per-token)	`RedHatAI/Qwen2.5-72B-Instruct-FP8-dynamic`	compressed-tensors fp8
mxfp8 (this repo)	`talumbau/Qwen2.5-72B-Instruct-mxfp8`	OCP MX FP8, block_size=32

License

This derivative work is distributed under the Qwen License Agreement.

Modifications: The original bfloat16 weights were post-training quantized to OCP MX FP8 format using llm-compressor v0.13.0. No fine-tuning or other modifications were made.