Property	Value
Format	FP8 E4M3 with block scaling
Weight block size	128×128
Activation group size	128 elements
Weight quantization	Static, per 128×128 block
Activation quantization	Dynamic, per 128-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

FP8 block acceleration requires hardware and a vLLM backend with block-scaled FP8 linear support. This checkpoint was validated on AMD MI350X (CDNA4, gfx950) with ROCm 7.2.3.

Note: Other hardware configurations have not been validated with this checkpoint.

Usage with vLLM

python

from vllm import LLM, SamplingParams

llm = LLM(
    model="rkayaith/Qwen2.5-72B-Instruct-fp8-block",
    # 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 rkayaith/Qwen2.5-72B-Instruct-fp8-block

Comparison variants

This checkpoint is part of a four-format study:

Variant	Repo	Format
BF16 (original)	`Qwen/Qwen2.5-72B-Instruct`	BF16
FP8 (per-channel/per-token)	`RedHatAI/Qwen2.5-72B-Instruct-FP8-dynamic`	Compressed-Tensors FP8
FP8 block (this repo)	`rkayaith/Qwen2.5-72B-Instruct-fp8-block`	FP8 E4M3, 128×128 weight blocks, dynamic 128-element activation groups
MXFP8	`talumbau/Qwen2.5-72B-Instruct-mxfp8`	OCP MXFP8, 32-element groups

License

This derivative work is distributed under the Qwen License Agreement.

Modifications: The original bfloat16 weights were post-training quantized to dynamic block FP8 format using LLM Compressor v0.13.1.dev2. No fine-tuning or other modifications were made.