festr2/GLM-5.2-BF16-KLD-Reference-Logits-20260708

GLM-5.2 BF16 KLD Reference Logits 20260708

This dataset contains the current GLM-5.2 BF16 reference prompt logits used for the July 2026 vLLM/Blackwell KLD checks.

The reference cache is intended for candidate-side KLD comparisons without rerunning the expensive BF16 reference pass.

Files

reference-logits/logits_0.safetensors
reference-logits/manifest.json
generation-log/config.env
generation-log/scoremode_kld.log

Reference Generation

Field	Value
Source model	zai-org/GLM-5.2
Snapshot	4d67f66cc64d3219133b767c253b2ad1425c6c88
Local source	/root/.cache/huggingface/hub/models--zai-org--GLM-5.2/snapshots/4d67f66cc64d3219133b767c253b2ad1425c6c88
Generated at	2026-07-08 00:04 UTC
Docker image	voipmonitor/vllm:eldritch-enlightenment-v7-vllme2e2eaf-b12x26144c0-cu132-20260707
vLLM version string	v0.11.2.dev279+eldritch.enlightenment.v7.vllme2e2eaf.b12x26144c0.cu132.20260707
Dataset	Salesforce/wikitext, wikitext-2-raw-v1, split test
Context length	2048
Stride	512
Windows	1
Tensor key	logits
Tensor shape	[2047, 154880]
Tensor dtype	float32
Attention backend	B12X_MLA_SPARSE
KV cache dtype during capture	fp8
Tensor parallel	16
DCP	1
Load format	safetensors

Tokenization first 16 ids:

[284, 8396, 425, 10960, 465, 284, 14721, 8396, 425, 10960, 465, 374, 458, 6364, 4531, 1154]

GLM-5.2 IndexCache override:

{"use_index_cache":true,"index_topk_pattern":"FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS"}

Download

mkdir -p /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z

huggingface-cli download festr2/GLM-5.2-BF16-KLD-Reference-Logits-20260708 \
  --repo-type dataset \
  --local-dir /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/hf

ln -sfn \
  /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/hf/reference-logits \
  /root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref

Inspect

python3 - <<'PY'
from safetensors import safe_open
path = "/root/kld/glm52_current_bf16_returnlogits_ref_20260708T000424Z/ref/logits_0.safetensors"
with safe_open(path, framework="pt", device="cpu") as f:
    for key in f.keys():
        tensor = f.get_tensor(key)
        print(key, tuple(tensor.shape), tensor.dtype)
PY

Expected:

logits (2047, 154880) torch.float32

The canonical reproduction page is:

https://github.com/local-inference-lab/rtx6kpro/blob/master/models/glm5.2/glm52-gguf-bf16-dequant-kld-2026-07-08.md