license: apache-2.0 library_name: transformers pipeline_tag: image-text-to-text base_model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored tags:

qwen
qwen3.5
qwen3.8
awq
int4
w4a16
compressed-tensors
lmdeploy
twin-turbo
heretic
uncensored
fable
davidau
finetune

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-W4A16-AWQ

W4A16 (4-bit weights, 16-bit activations) AWQ compressed-tensors quantization of DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored.

TWIN-TURBO: Smaller quants with higher performance and vastly reduced "thinking tokens". Features 5 thinking modes and 5 instruct modes, switchable on the fly (via API, direct, or chat "in message"). This is the Light to Moderate Heretic/uncensored version with stronger balance on performance.

The source model has since shipped updated jinja chat templates (fixing tool-calling); chat_template.jinja here is the current version from the source repo.

Quantization method

Scheme: W4A16_ASYM — 4-bit asymmetric per-group quantization (group size 128) of all Linear weights, stored in the compressed-tensors pack-quantized format (weight_packed / weight_scale / weight_zero_point / weight_shape), which LMDeploy turbomind auto-detects and loads natively (including the MTP heads and vision tower, which stay BF16).
Tooling: llmcompressor one-shot offline quantization with CPU offloading (compressed_tensors.offload.load_offloaded_model), so the full-precision source fits on a 2×16 GB VRAM setup.
AWQ activation smoothing: AWQModifier with the layer-scoped hybrid-attention mappings from build_hybrid_attention_mappings — full-attention input_layernorm → self_attn.q/k/v, post_attention_layernorm → mlp.gate/up, and mlp.up_proj → mlp.down_proj, with duo_scaling="both" and CPU offload, followed by W4A16 quantization. This layer-scoped recipe is required for hybrid-attention (Qwen3.5-family) architectures — grouped-regex smoothing or mismatched mappings corrupt decoding.
Unquantized (kept BF16): embeddings, lm_head, norms, linear_attn.in_proj_a/b, the vision tower, and MTP heads.
Quant run stats: 244.0 min total, peak 5.06 GB per GPU. 400 quantized modules across 7 data shards + nonquant shard.

Benchmarks

See the original model's README for full benchmark tables.

Usage

Tested with LMDeploy turbomind:

from lmdeploy import pipeline, TurbomindEngineConfig

pipe = pipeline(
    "TheUnderscore/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-W4A16-AWQ",
    backend_config=TurbomindEngineConfig(
        tp=2,
        model_format="compressed-tensors",
        language_model_only=True,
    ),
)
print(pipe("Hello, who are you?").text)

Files

chat_template.jinja — the current chat template from the source repo. Also included: chat_template-old.jinja, chat_template-toolcall2.jinja, and the larger chat_template-tturbo.jinja (enhanced tool-calling) variant. None are referenced in tokenizer_config.json, so the loader uses its own default. To switch, swap the contents of chat_template.jinja with a variant (or point your server's chat-template option at the file).
quantize-awq-hybrid.py — the script used to produce this quantization (CPU-offloaded, DDP/torchrun, produces properly numbered -of-N shards). --offload_dir selects where per-rank CPU offload temp folders live (defaults to the current working directory).
model-nonquant.safetensors — unquantized tensors (mtp.* and model.visual.*) preserved BF16 so the full model architecture is loadable.

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-W4A16-AWQ

The source model has since shipped updated jinja chat templates (fixing tool-calling); chat_template.jinja here is the current version from the source repo.

Quantization method

Scheme: W4A16_ASYM — 4-bit asymmetric per-group quantization (group size 128) of all Linear weights, stored in the compressed-tensors pack-quantized format (weight_packed / weight_scale / weight_zero_point / weight_shape), which LMDeploy turbomind auto-detects and loads natively (including the MTP heads and vision tower, which stay BF16).

Tooling: llmcompressor one-shot offline quantization with CPU offloading (compressed_tensors.offload.load_offloaded_model), so the full-precision source fits on a 2×16 GB VRAM setup.

AWQ activation smoothing: AWQModifier with the layer-scoped hybrid-attention mappings from build_hybrid_attention_mappings — full-attention input_layernorm → self_attn.q/k/v, post_attention_layernorm → mlp.gate/up, and mlp.up_proj → mlp.down_proj, with duo_scaling="both" and CPU offload, followed by W4A16 quantization. This layer-scoped recipe is required for hybrid-attention (Qwen3.5-family) architectures — grouped-regex smoothing or mismatched mappings corrupt decoding.

Unquantized (kept BF16): embeddings, lm_head, norms, linear_attn.in_proj_a/b, the vision tower, and MTP heads.

Quant run stats: 244.0 min total, peak 5.06 GB per GPU. 400 quantized modules across 7 data shards + nonquant shard.

Usage

Tested with LMDeploy turbomind:

from lmdeploy import pipeline, TurbomindEngineConfig pipe = pipeline( "TheUnderscore/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-W4A16-AWQ", backend_config=TurbomindEngineConfig( tp=2, model_format="compressed-tensors", language_model_only=True, ), ) print(pipe("Hello, who are you?").text)

Files

chat_template.jinja — the current chat template from the source repo. Also included: chat_template-old.jinja, chat_template-toolcall2.jinja, and the larger chat_template-tturbo.jinja (enhanced tool-calling) variant. None are referenced in tokenizer_config.json, so the loader uses its own default. To switch, swap the contents of chat_template.jinja with a variant (or point your server's chat-template option at the file).

quantize-awq-hybrid.py — the script used to produce this quantization (CPU-offloaded, DDP/torchrun, produces properly numbered -of-N shards). --offload_dir selects where per-rank CPU offload temp folders live (defaults to the current working directory).

model-nonquant.safetensors — unquantized tensors (mtp.* and model.visual.*) preserved BF16 so the full model architecture is loadable.