Qwen3.5-122B-A10B-heretic-v2-FP8

FP8 block-wise quantization of coder3101/Qwen3.5-122B-A10B-heretic-v2 (abliterated via Heretic).

Quantization format matches Qwen/Qwen3.5-122B-A10B-FP8 exactly (same tensor names, scale format, skip list).

Quantization Details

Parameter	Value
Method	FP8 E4M3 block-wise
Block size	128x128
Scale dtype	BF16
Scale naming	`weight_scale_inv`
Activation scheme	Dynamic
Original size	~227 GB (BF16)
Quantized size	~117 GB

What's quantized

All linear projections in attention (q/k/v/o_proj), GatedDeltaNet (in_proj_qkv, in_proj_z, out_proj), and MLP experts (gate/up/down_proj for all experts + shared expert).

What's kept in BF16

Embeddings, lm_head, norms, MoE router gates, GDN precision-sensitive params (conv1d, in_proj_a, in_proj_b), vision tower.

Usage with SGLang (2+ GPUs)

python -m sglang.launch_server \
    --model-path nivvis/Qwen3.5-122B-A10B-heretic-v2-FP8 \
    --tp 2 \
    --port 30000 --host 0.0.0.0 \
    --mem-fraction-static 0.85 \
    --context-length 32768 \
    --attention-backend triton \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder \
    --mamba-scheduler-strategy extra_buffer \
    --trust-remote-code

Usage with vLLM

from vllm import LLM
model = LLM("nivvis/Qwen3.5-122B-A10B-heretic-v2-FP8", tensor_parallel_size=2, trust_remote_code=True)

Notes

No MTP: Heretic abliteration strips MTP weights. Do not use --speculative-algo NEXTN (accept rate ~0.25). Restoring MTP by copying the head from the original Qwen model and fine-tuning against heretic hidden states is a potential future improvement.
Thinking model: Supports <think> reasoning. Use "chat_template_kwargs": {"enable_thinking": false} to disable.
VRAM: ~117 GB model weights. Requires 2x H100 80GB, 1x H200 141GB (tight), or 2x RTX PRO 6000 96GB.
Validated against official Qwen/Qwen3.5-122B-A10B-FP8: 0 missing scale tensors, 0 unexpected tensors, matching quantization config.