m8than/GLM-5.2-E2M1-Furnace

GLM-5.2 E2M1 for Furnace

Complete experimental conversion of zai-org/GLM-5.2-FP8 for Furnace. 142 safetensors shards, 117,953 tensors, approximately 405 GB. The source checkpoint is unchanged. This is a full target model, not a pilot or draft.

Serving

Measured hardware: eight AMD MI325X GPUs (gfx942); ROCm 7.2, PyTorch 2.11, Triton 3.6, AITER revision c16d44b93a528b2a4bfd6d8d3409116d465872a9. See Furnace's START_COMMANDS.md for the full TP8 launch recipe. Set:

--model-path m8than/GLM-5.2-E2M1-Furnace \
--tp-size 8 --dtype bfloat16 --kv-cache-dtype bfloat16 \
--attention-backend dsa --dsa-prefill-backend aiter --dsa-decode-backend aiter

The model can use the separately published experimental DFlash2 B12 conversation refresh. That draft's acceptance improved on one held-out conversation test, but it did not consistently improve speed over the preceding local conversation draft. Draft weights are not included in this repository.

--superfast is a separate, opt-in lossy prefill policy, not a requirement of the checkpoint format. Check the pinned Furnace revision's plan and launch documentation for its attention and prefix-cache behavior. Do not use lossy superfast serving for exact-teacher capture.

Checks and limitations

Conversion checks verified the source against the pinned Hugging Face tree, all 142 output-shard checksums, complete tensor/index coverage, and tokenizer assets. Publication checks remote shard sizes and SHA256 against conversion receipts. conversion_metadata.json contains the public aggregate provenance and shard checksums; no raw conversations, caches or local paths are included.

Before publication, the retained target/runtime combination with the previous conversation draft scored 98.125% on a fixed, finite 160-task suite, matching the unchanged normal FP8 baseline's point score. This does not establish universal equivalence: the suite is templated, writing grading was not independent human evaluation, and one long-context instruction task regressed. The test combined E2M1 with lossy superfast attention; it does not isolate the quantization effect. No new broad quality run covered the refreshed draft.

On the then-retained runtime, a single-stream 32,768-input/256-output-token test with the previous draft measured median decode 169.1 tok/s and effective prefill 20,512.6 input tok/s (input tokens / time to first token). Three trials followed one excluded warmup. The refresh measured 175.8 and 20,502.5 tok/s at 32K, but regressed at shorter contexts. 250 decode tok/s has not been achieved consistently. These are workload-specific observations, not a clean-machine deployment guarantee. Configured maximum positions do not prove long-context quality or available KV capacity.

The original MIT license is included unchanged. The separate draft has its own inherited donor-license limitations; this target's MIT license does not license the draft weights.