Downloads · 30 days
10
19% of all-time downloads
tchedaTT/Qwen3.8-27B-emax-minus-one-bf16
Qwen3.8-27B-emax-minus-one-bf16 is a image-text-to-text model from tchedaTT. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
This is an experimental derivative of Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Selected weights were modified on 2026-09-21 using offline, weight-only exponent search for Tenstorrent BFP4B…
Downloads · 30 days
10
19% of all-time downloads
All-time downloads
52
Public
Parameters
27.8B
55.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors55.6 GB · 100%
From the Hugging Face model README
This is an experimental derivative of Qwen/Qwen3.8-27B, revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Selected weights were modified on
2026-09-21 using offline, weight-only exponent search for Tenstorrent BFP4_B
and BFP8_B. The files remain ordinary BF16 safetensors, approximately 55.6 GB.
This is not a compressed INT4 checkpoint or an upstream Qwen release.
Prepared for mando2222/qwen3.8-27b-dflash2-p300x2-q4kv at revision
9a39bc16ca0a9d560f06caa2111db6c25fa0b524, using its existing dtypes and
4-chip tensor-parallel layout. The bundle's q4kv setting is about the KV
cache and is unchanged. Its DFlash2 draft checkpoint is also unchanged.
For each group of 16 outputs per input channel, evaluate the ordinary maximum exponent and maximum exponent minus one, quantize with TT mantissa rounding and saturation, and keep the candidate with smaller unweighted squared weight error. Ties keep the ordinary exponent. Store the resulting representable values in BF16. There is no calibration, inference, training, GPTQ, or activation scaling.
Modified matrices:
| Target text-model weights | Bundle format (unchanged) |
|---|---|
| MLP gate and up | BFP4_B |
| MLP down | BFP8_B |
| Full-attention Q, K, V, O | BFP8_B |
| GDN QKV, Z and output projections | BFP8_B |
| Language-model output head | BFP8_B |
Embeddings, norms, convolutions, small GDN A/B projections, vision parameters,
and MTP parameters are unchanged. The small GDN A/B projections are excluded
because their fused per-device widths cross 16-element exponent boundaries.
The source configuration, tokenizer, tensor names, shapes and BF16 storage
dtypes are preserved. tt_bfp_quantization.json records every modified tensor,
source/output hashes, group choices, reconstruction errors and validation.
Every changed group is checked for no increase in squared weight error versus ordinary quantization. Every exported value is exactly representable in BF16 and survives ordinary max-exponent re-quantization for the audited grouping. Every serialized tensor is hash-checked; untargeted tensors match the source.
No full-model TT inference, perplexity, task accuracy, or performance benchmark has been run on this checkpoint. Lower weight error does not establish improved model accuracy. BF16 is a convenient transport approximation: changing runtime dtypes, grouping, sharding, fusions or quantization code requires revalidation. Use a fresh TT tensor cache when switching checkpoints.
On the serving host, authenticate to Hugging Face if this repository is private.
Use the same HF_HOME / HF_HUB_CACHE for all these commands. Download
prepare_bundle.py from this repository, then substitute this repository's ID
and exact commit below:
export CHECKPOINT_REPO='YOUR_NAMESPACE/Qwen3.8-27B-emax-minus-one-bf16'
export CHECKPOINT_REV='EXACT_40_CHARACTER_COMMIT'
tt-model pull mando2222/qwen3.8-27b-dflash2-p300x2-q4kv@9a39bc16ca0a9d560f06caa2111db6c25fa0b524
hf download "$CHECKPOINT_REPO" --revision "$CHECKPOINT_REV"
hf download incoai/Qwen3.8-27B-DFlash2 --revision dedf8df68adfb1afeaf7b7480c0a0243108177b4
python prepare_bundle.py "$CHECKPOINT_REPO" --revision "$CHECKPOINT_REV" --out qwen38-emaxm1-bundle
tt-model serve --print qwen38-emaxm1-bundle/tt_kernel_manifest.json
tt-model serve qwen38-emaxm1-bundle/tt_kernel_manifest.json
The helper creates a local manifest with a separate cache name and exact snapshot selection. It preserves the image, model code, runtime precision, KV-cache format and DFlash2 checkpoint. It does not upload or start a server.
export_weights.py is the bounded-memory export driver. It uses the offline
tt_bfp_quant package from tt-metal branch tcheda/offline-bfp-quantization,
commit 75052cb05887623a82e173b2a46d3d6b4b78ab45, with exponent candidates
[0, -1]. The adjacent audit manifest records the pinned source and scope.
The original Apache-2.0 LICENSE and original model card (UPSTREAM_README.md)
are retained. Copyright 2026 Alibaba Cloud is retained in the license.
The original card describes the upstream model; its published evaluation
results must not be interpreted as measurements of this modified checkpoint.
Upstream crc32.txt is omitted because those checksums no longer describe the
changed shards. Use SHA256SUMS for this export.