soyrsoyr/GLM-5.3-Flash-NVFP4A16-MTP-PR3118

GLM-5.3-Flash NVFP4A16 with MTP — PR3118 validation

Full pretrained checkpoint converted from zai-org/GLM-5.3-Flash using llm-compressor PR3118, `7029daba`.

Checkpoint structure and MTP tensor checks passed. B200 MTP serving passed: four HTTP completions, positive draft-token metrics, and successful worker cleanup. See b200-runtime.json for generations, acceptance counters, and the exact tested revision. This is a serving smoke test using raw completion prompts without a chat template, not a model-quality benchmark. No quality or performance benchmark is claimed.

The backbone MLP gate/up/down projections and supported MTP MLP projections use data-free NVFP4A16: weight-only FP4 with 16-bit activations. Other projections remain BF16. Native FP8 source weights are dequantized before conversion. See recipe.yaml, config.json, and pr3118-validation.json for details.

The output generation defaults enable sampling to make the source top_p=0.95 setting valid under Transformers 5.17. The runtime smoke test explicitly uses temperature 0.

The source model's licensing applies; consult its model card and included license files.

Validated runtime: vLLM 0.29.1rc1.dev79+g767d1c4d4, two B200 GPUs, one speculative MTP token, block size 256. This artifact is intended for PR validation.

B200 smoke-test metrics

See runtime evidence for the actual generations and counters. The acceptance ratio is from this four-request smoke test, not a quality or throughput benchmark.