Datasetpublic

festr2/glm51-kld-reference-logits-wikitext-ctx2048-s512-20260517

--- license: other pretty_name: GLM-5.1 KLD Reference Logits WikiText ctx2048 s512 tags: - logits - kld - glm-5.1 - vllm - b12x --- # GLM-5.1 KLD Reference Logits Public cache of the reference logits used for GLM-5.1 NVFP4 / mixed FP8_PB_WO KLD evaluation. These files are generated logits, not model weights. They are stored as `logits_*.safetensors` with one tensor named `logits`, shape `(2047, 154880)`, dtype `float32`. ##… See the full description on the dataset page: https://huggingface.co/datasets/festr2/glm51-kld-reference-logits-wikitext-ctx2048-s512-20260517.

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes129downloads

Dataset Card

license: other pretty_name: GLM-5.1 KLD Reference Logits WikiText ctx2048 s512 tags:

—logits
—kld
—glm-5.1
—vllm
—b12x ---

# GLM-5.1 KLD Reference Logits

Public cache of the reference logits used for GLM-5.1 NVFP4 / mixed FP8PBWO KLD evaluation. These files are generated logits, not model weights. They are stored as logits_*.safetensors with one tensor named logits, shape (2047, 154880), dtype float32.

## Contents

path	kind	windows	files	size	source checkpoint
fp8-b12xmlasparse-w1	FP8 reference logits	1	1	1.18 GiB	zai-org/GLM-5.1-FP8@a92f8155fe8574288534d41a6dbc1d72888ab2da
fp8-b12xmlasparse-w8	FP8 reference logits	8	8	9.45 GiB	zai-org/GLM-5.1-FP8@a92f8155fe8574288534d41a6dbc1d72888ab2da
fp8-b12xmlasparse-w42	FP8 reference logits	42	42	49.60 GiB	zai-org/GLM-5.1-FP8@a92f8155fe8574288534d41a6dbc1d72888ab2da
bf16-b12xmlasparse-w1	BF16 reference logits	1	1	1.18 GiB	zai-org/GLM-5.1@26e1bd6e011feb778d25ae34b09b07074139d92d

## Generation Setup

—Dataset: wikitext, config wikitext-2-raw-v1
—Context length: 2048
—Stride: 512
—Tensor parallel size: 8
—Attention backend: B12X_MLA_SPARSE
—KV cache dtype during generation: fp8
—GLM index pattern: FFSFSSSFSSFFFSSSFFFSFSSSSSSFFSFFSFFSSFFFFFFSFFFFFSFFSSSSSSFSFFFSFSSSFSFFSFFSSS
—Runner image: voipmonitor/vllm:glm51-canonical-git-vllmcf2070d-b12xc929144-cutedsl45-20260513
—vLLM branch: voipmonitor/vllm@codex/glm51-kimi-canonical-rebase-test-20260514

The BF16 W1 reference was generated from zai-org/GLM-5.1 snapshot 26e1bd6e011feb778d25ae34b09b07074139d92d with CPU offload because the full BF16 checkpoint does not fit on 8x RTX PRO 6000 Blackwell.

## Usage

Point the vLLM KLD runner at one of these directories with --reference-logits, for example:

bash

        python3 examples/offline_inference/score_mode_kld.py \
          --model /path/to/candidate-checkpoint \
          --reference-logits /path/to/fp8-b12xmlasparse-w42 \
          --dataset wikitext --dataset-config wikitext-2-raw-v1 \
          --tensor-parallel-size 8 \
          --attention-backend B12X_MLA_SPARSE

Logs used to generate the artifacts are under logs/. MANIFEST.json records local source paths, file counts, source snapshots, and known legacy references that were intentionally not uploaded.