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, 118,629 tensors, approximately 388 GB. The source checkpoint is unchanged. This is a full target model, not a pilot or draft.

What changed in this revision

Routed experts are E2M1 (4-bit, block-128 scales), as before. Attention, dense and shared-expert projections are now FP8 block-128, passed through unchanged from the source (dense_weight_format: fp8_block128 in quantization_config), instead of BF16. Expert weights and their reconstruction error are identical to the previous revision. The previous BF16-dense revision is 16c3f5ca406349306ffe69de7726d7156b09b91b.

This needs a Furnace build that understands dense_weight_format (the commit linked above or later). FP8 dense weights let Furnace's FURNACE_MEGAKERNEL=1 decode path qualify on this checkpoint.

Serving

Measured hardware: four AMD MI325X GPUs (gfx942), TP4, native MTP (EAGLE 5/4/9), --superfast fast. See Furnace's START_COMMANDS.md for the full launch recipe.

--model-path m8than/GLM-5.2-E2M1-Furnace \
--tp-size 4 --dtype bfloat16 --kv-cache-dtype bfloat16 \
--attention-backend dsa --dsa-prefill-backend aiter --dsa-decode-backend aiter

--superfast is a separate, opt-in lossy prefill policy, not a requirement of the checkpoint format. FURNACE_MEGAKERNEL is an experimental, off-by-default decode path; it qualifies only at TP4 with FP8 block-128 dense weights.

Checks and limitations

Conversion verified the source against the pinned Hugging Face tree, every output-shard checksum, complete tensor/index coverage and the tokenizer assets. conversion_metadata.json holds the aggregate provenance and shard checksums; no conversations, caches or local paths are included.

Before publication, one interleaved A/B round on 4x MI325X (TP4, EAGLE 5/4/9, --superfast fast), stock vs FURNACE_MEGAKERNEL=1, single run per arm:

Decode tok/s	stock	FURNACE_MEGAKERNEL=1
B1	151.6	181.4
B4 (total)	489.2	535.7
B8 (total)	795.5	873.5
B16 (total)	892.5	910.7
B24 (total)	1116.4	1157.2

Prefill was unchanged (32K: 9,898 vs 9,980 tok/s). Furnace's quick check (sanity, 12K/32K needles, 40-key passkey, multi-turn, GSM8K subset) passed on both arms. These are workload-specific single-round observations on a shared machine, not a deployment guarantee; KV capacity measured 1,374,144 tokens at TP4 with --mem-fraction-static 0.85 (stock path). Configured maximum positions do not prove long-context quality.

The original MIT license is included unchanged.