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.
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.
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.
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.
model-source-00000-0000.safetensors 4.3 GBmodel-source-00019-0000.safetensors 3.0 GBmodel-source-00135-0000.safetensors 2.9 GBmodel-source-00137-0000.safetensors 2.8 GBmodel-source-00010-0000.safetensors 2.8 GBmodel-source-00056-0000.safetensors 2.8 GBmodel-source-00032-0000.safetensors 2.8 GBmodel-source-00113-0000.safetensors 2.8 GBmodel-source-00089-0000.safetensors 2.8 GBmodel-source-00065-0000.safetensors 2.8 GBmodel-source-00017-0000.safetensors 2.8 GBmodel-source-00041-0000.safetensors 2.8 GBmodel-source-00122-0000.safetensors 2.8 GBmodel-source-00100-0000.safetensors 2.8 GBmodel-source-00002-0000.safetensors 2.8 GBmodel-source-00024-0000.safetensors 2.8 GBmodel-source-00048-0000.safetensors 2.8 GBmodel-source-00072-0000.safetensors 2.8 GBmodel-source-00081-0000.safetensors 2.8 GBmodel-source-00105-0000.safetensors 2.8 GBmodel-source-00129-0000.safetensors 2.8 GBmodel-source-00096-0000.safetensors 2.8 GBmodel-source-00069-0000.safetensors 2.8 GBmodel-source-00034-0000.safetensors 2.8 GBmodel-source-00115-0000.safetensors 2.8 GBmodel-source-00045-0000.safetensors 2.8 GBmodel-source-00126-0000.safetensors 2.8 GBmodel-source-00091-0000.safetensors 2.8 GBmodel-source-00021-0000.safetensors 2.8 GBmodel-source-00080-0000.safetensors 2.8 GBmodel-source-00067-0000.safetensors 2.8 GBmodel-source-00102-0000.safetensors 2.8 GBmodel-source-00008-0000.safetensors 2.8 GBmodel-source-00124-0000.safetensors 2.8 GBmodel-source-00043-0000.safetensors 2.8 GBmodel-source-00054-0000.safetensors 2.8 GBmodel-source-00006-0000.safetensors 2.8 GBmodel-source-00030-0000.safetensors 2.8 GBmodel-source-00111-0000.safetensors 2.8 GBmodel-source-00076-0000.safetensors 2.8 GBmodel-source-00087-0000.safetensors 2.8 GBmodel-source-00052-0000.safetensors 2.8 GBmodel-source-00133-0000.safetensors 2.8 GBmodel-source-00063-0000.safetensors 2.8 GBmodel-source-00004-0000.safetensors 2.8 GBmodel-source-00028-0000.safetensors 2.8 GBmodel-source-00109-0000.safetensors 2.8 GBmodel-source-00074-0000.safetensors 2.8 GBmodel-source-00015-0000.safetensors 2.8 GBmodel-source-00046-0000.safetensors 2.8 GBmodel-source-00050-0000.safetensors 2.8 GBmodel-source-00085-0000.safetensors 2.8 GBmodel-source-00011-0000.safetensors 2.8 GBmodel-source-00026-0000.safetensors 2.8 GBmodel-source-00035-0000.safetensors 2.8 GBmodel-source-00061-0000.safetensors 2.8 GBmodel-source-00083-0000.safetensors 2.8 GBmodel-source-00092-0000.safetensors 2.8 GBmodel-source-00094-0000.safetensors 2.8 GBmodel-source-00107-0000.safetensors 2.8 GBmodel-source-00116-0000.safetensors 2.8 GBmodel-source-00127-0000.safetensors 2.8 GBmodel-source-00131-0000.safetensors 2.8 GBmodel-source-00013-0000.safetensors 2.8 GBmodel-source-00070-0000.safetensors 2.8 GBmodel-source-00037-0000.safetensors 2.8 GBmodel-source-00118-0000.safetensors 2.8 GBmodel-source-00139-0000.safetensors 2.8 GBmodel-source-00078-0000.safetensors 2.8 GBmodel-source-00098-0000.safetensors 2.8 GBmodel-source-00039-0000.safetensors 2.8 GBmodel-source-00120-0000.safetensors 2.8 GBmodel-source-00059-0000.safetensors 2.8 GBmodel-source-00104-0000.safetensors 2.8 GBmodel-source-00057-0000.safetensors 2.8 GBmodel-source-00023-0000.safetensors 2.8 GBmodel-source-00022-0000.safetensors 2.7 GBmodel-source-00058-0000.safetensors 2.7 GBmodel-source-00103-0000.safetensors 2.7 GBmodel-source-00001-0000.safetensors 2.7 GBmodel-source-00012-0000.safetensors 2.7 GBmodel-source-00025-0000.safetensors 2.7 GBmodel-source-00036-0000.safetensors 2.7 GBmodel-source-00047-0000.safetensors 2.7 GBmodel-source-00049-0000.safetensors 2.7 GBmodel-source-00060-0000.safetensors 2.7 GBmodel-source-00071-0000.safetensors 2.7 GBmodel-source-00082-0000.safetensors 2.7 GBmodel-source-00084-0000.safetensors 2.7 GBmodel-source-00093-0000.safetensors 2.7 GBmodel-source-00095-0000.safetensors 2.7 GBmodel-source-00106-0000.safetensors 2.7 GBmodel-source-00117-0000.safetensors 2.7 GBmodel-source-00128-0000.safetensors 2.7 GBmodel-source-00130-0000.safetensors 2.7 GBmodel-source-00014-0000.safetensors 2.7 GBmodel-source-00073-0000.safetensors 2.7 GBmodel-source-00027-0000.safetensors 2.7 GBmodel-source-00108-0000.safetensors 2.7 GBmodel-source-00003-0000.safetensors 2.7 GBmodel-source-00062-0000.safetensors 2.7 GBmodel-source-00097-0000.safetensors 2.7 GBmodel-source-00132-0000.safetensors 2.7 GBmodel-source-00051-0000.safetensors 2.7 GBmodel-source-00086-0000.safetensors 2.7 GBmodel-source-00040-0000.safetensors 2.7 GBmodel-source-00121-0000.safetensors 2.7 GBmodel-source-00016-0000.safetensors 2.7 GBmodel-source-00075-0000.safetensors 2.7 GBmodel-source-00029-0000.safetensors 2.7 GBmodel-source-00110-0000.safetensors 2.7 GBmodel-source-00005-0000.safetensors 2.7 GBmodel-source-00064-0000.safetensors 2.7 GBmodel-source-00134-0000.safetensors 2.7 GBmodel-source-00053-0000.safetensors 2.7 GBmodel-source-00088-0000.safetensors 2.7 GBmodel-source-00042-0000.safetensors 2.7 GBmodel-source-00123-0000.safetensors 2.7 GBmodel-source-00018-0000.safetensors 2.7 GBmodel-source-00077-0000.safetensors 2.7 GBmodel-source-00031-0000.safetensors 2.7 GBmodel-source-00112-0000.safetensors 2.7 GBmodel-source-00007-0000.safetensors 2.7 GBmodel-source-00066-0000.safetensors 2.7 GBmodel-source-00136-0000.safetensors 2.7 GBmodel-source-00101-0000.safetensors 2.7 GBmodel-source-00020-0000.safetensors 2.7 GBmodel-source-00055-0000.safetensors 2.7 GBmodel-source-00090-0000.safetensors 2.7 GBmodel-source-00125-0000.safetensors 2.7 GBmodel-source-00044-0000.safetensors 2.7 GBmodel-source-00033-0000.safetensors 2.7 GBmodel-source-00114-0000.safetensors 2.7 GBmodel-source-00009-0000.safetensors 2.7 GBmodel-source-00068-0000.safetensors 2.7 GBmodel-source-00038-0000.safetensors 2.7 GBmodel-source-00119-0000.safetensors 2.7 GBmodel-source-00099-0000.safetensors 2.7 GBmodel-source-00079-0000.safetensors 2.7 GBmodel-source-00138-0000.safetensors 2.7 GBmodel-source-00140-0000.safetensors 2.5 GBmodel-source-00000-0001.safetensors 703 MB有文件?提交种子。
提交种子zai-org/GLM-5.2-FP8