Launch80 Swift-1.5 Qwen3.8-27B Uncensored — INT4 GPTQ + 12 fused-qkv BF16 groups + L63 o_proj BF16

The flagship mixed-precision checkpoint of the Launch80 quality pipeline: the Launch80 INT4 bake with 12 Hessian-selected full-attention groups promoted to BF16 fused qkv_proj tensors plus the layer-63 self_attn.o_proj in BF16. This is the checkpoint serving production on the author's stack since 2026-10-27.

Provenance

Derived from davetha/launch80-swift15-qwen27b-unc-int4-bodyp-bf16 (INT4 + L63 o_proj BF16). The 12 promoted groups (layers 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63 self_attn qkv) were selected by activation-Hessian capture + activation-space error scoring on the live INT4 target, then fused into single BF16 qkv_proj.weight tensors (tree-form names: language_model.model.layers.N.self_attn.qkv_proj.weight). Packaging carries modules_in_block_to_quantize and is stripped of the orphaned INT4 q/k/v tensors the fusion supersedes (144 removed), so stock loaders never see stale shards.

Serving recipe (int8 weight islands on a native int8 GEMM)

The intended serving mode — and what production runs — replaces the 13 BF16 islands with resident per-channel symmetric int8 + a native oneDNN int8 GEMM at load time:

• vLLM 0.30 XPU stack + a vllm-xpu-kernels build carrying int8_gemm_w8a16 (oneDNN W8A16 primitive; per-channel scales via group_size=K — verified bit-exact against bf16 dequant).
• Env: K8V4_INT8_ISLANDS=layers.19.self_attn.qkv_proj,...,layers.63.self_attn.qkv_proj,layers.63.self_attn.o_proj (13 modules; islands quantize on load — this checkpoint stays BF16 on disk).
• DFlash2 speculative drafter (k=7), fp8 KV cache, --max-model-len 131072.
• Loading the fused BF16 qkv tensors on stock vLLM ≤0.30 needs the loader fix vllm-project/vllm#63631 (name-form variants + anchor-safe dynamic matching; proven end-to-end on this checkpoint family, including a full-context CUDA validation).

Measured quality (task-accuracy eval, 3 seeds, temp 0)

GSM8K-400 / MMLU-500 / HumanEval-60 prefix-EM, seeds {0,1,2}:

Decode speed at code-generation parity with the BF16-island arm (~139 tok/s single-stream on a B70-class 32 GB card with spec decode); int8 islands halve island VRAM vs BF16.

Limitations

• The 12-group promotion trades ~0.8 pp MMLU vs the unfused sibling for +5 pp GSM8K — a deliberate reasoning-vs-knowledge profile, not a strict improvement.
• The int8-island path requires the patched kernel stack described above; without it, serve with BF16 islands directly (this checkpoint loads either way).
• GPTQ INT4 artifacts apply; behavior profile inherits the uncensored base.