Modelpublic

Mantisec/Qwen3.8-27B-OBLITERATED-FP16

sourceHugging Faceupdated 25d agoView on Hugging Face

0likes95downloads

Model Card

Qwen3.8-27B-OBLITERATED-fp16

FP16 conversion of `OBLITERATUS/Qwen3.8-27B-OBLITERATED` at revision a58c3b53b3ce71551eafde2ed5ec8df48e0f4ff8, produced by bfsquish v0.1.0.

Intended use

Optimized for FP16 inference and fine-tuning on NVIDIA V100 (Volta) GPUs, which lack native BF16 Tensor Core support. The weights have been converted from BF16 to FP16 using the effective range_checked strategy (see below).

Conversion details

Field	Value
Source model	`OBLITERATUS/Qwen3.8-27B-OBLITERATED`
Source revision	`a58c3b53b3ce71551eafde2ed5ec8df48e0f4ff8`
Requested strategy	`auto`
Effective strategy	`range_checked`
Tool	`bfsquish` v0.1.0
Converted at (UTC)	`2026-08-31T18:46:46.788873+00:00`
Target hardware	NVIDIA V100 (Volta, sm_70)
Target runtime	NVIDIA V100 (Volta sm_70), FP16 Tensor Cores

Conversion strategy

Direct upcast to FP32 followed by a range-checked FP16 cast. Values outside FP16's finite range are rejected instead of clipped, and rounding error is recorded in bfsquish_conversion.json. For BF16-trained weights that already fit inside FP16's range, this avoids graph-changing rotations and preserves the source checkpoint as closely as the target dtype permits.

Transform plan

v1 strategy applied: range_checked (no v2 plan bundle was produced for this conversion).

Numerical quality

Metric	Value
Validation verdict	PASS
Validation method	`generate`
Max abs logit diff (vs. source)	0.157488
Min cosine similarity (vs. source)	0.999983
Token agreement rate	100.00%
Inf/NaN scan	passed (no inf/nan)

Validation notes

—generation agreement passed despite logit drift: tokenagreement=100.00% (pass≥98.00%), mincos=0.999983, max_diff=0.1575

Reproducing this conversion

bash

bfsquish run \
  --model OBLITERATUS/Qwen3.8-27B-OBLITERATED \
  --output-dir ./out \
  --strategy auto

License

Inherited from the source model. Refer to the source model's license for terms of use.