transformers 5 stores routed experts as fused 3D parameters
(experts.gate_up_proj / experts.down_proj) for Mixtral, Qwen3-MoE,
DeepSeek-V3, GLM-4 MoE, Llama 4, gpt-oss and OLMoE. When such a layer is
CPU- or disk-offloaded by Accelerate, EXCISE failed closed with
"offloaded fused expert tensors are not yet a supported surgery layout",
so every frontier MoE checkpoint large enough to need offload could not be
abliterated at all.
Route the three fused projections and the fused bias projection through the
existing LogicalParameterTransaction adapter: a meta-resident fused
parameter is resolved to a private copy of its authoritative backing value,
projected per expert, and committed atomically; the live parameter stays on
meta and quantized or unknown backing layouts still fail before mutation.
Thread offload_roots through the MoE dispatch call sites so parent-prefixed
Accelerate hooks resolve, and drop the preflight rejection of 3D meta
parameters (validate_offloaded_parameters already resolves them).
Tests: fused 3D, bias, granular and selective-inversion projections on
offloaded backing stores, bounded norm restoration, parent-prefixed hook
resolution, quantized fail-closed, commit-failure rollback, and a full
offline pipeline run on a disk-offloaded tiny Mixtral fixture that verifies
the saved checkpoint carries the updated experts with no meta tensors.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Carry the coherent blending contribution and research summary from PR #127 while splitting the capacity and recovery proposals into issues #132 and #133.
Point obliteratus at an FP8 or NVFP4 checkpoint and it just works:
the loader detects the format from config.json + safetensors metadata
(no weight loads), dequantizes shard-by-shard to a temporary BF16 copy,
then runs the normal float pipeline and saves BF16.
Supported layouts:
- FP8 DeepSeek-style block-wise (weight_scale_inv + weight_block_size)
- FP8 per-channel / per-tensor (compressed-tensors, ModelOpt)
- NVFP4 ModelOpt (uint8 nibbles + FP8 group scales + FP32 global),
including MIXED_PRECISION checkpoints (FP8 mixer + NVFP4 experts)
- NVFP4 compressed-tensors (reciprocal scales)
Design:
- New pure-torch obliteratus/models/quant_dequant.py; no new deps.
NVFP4 unpack uses torch.float4_e2m1fn_x2 when a runtime probe proves
it works, else a chunked nibble LUT (bounds transient int64 index
memory; a naive implementation OOMed at 96GB on a 30B model).
- Scale keys are dropped only when their base weight exists in the
same shard, so legitimate params ending in _scale (logit_scale et al.)
survive.
- Unsupported schemes (fbgemm, quanto, W4A4, ...) fail loudly at load,
naming the scheme.
- Surgery guards: float8 or packed uint8 reaching _dequantize_weight or
any fused-MoE path raises RuntimeError instead of silently upcasting
(bitsandbytes quant_state params are explicitly excluded).
- Save path strips quantization metadata and logs that output is BF16;
re-quantization for serving is out of scope (llm-compressor/modelopt).
- CLI: new --trust-remote-code flag; help text documents auto-detection.
Validated end-to-end on 1x A100-80GB (see PR description):
Nemotron-3-Nano-Omni-30B NVFP4 (mixed) and FP8, Qwen3-8B-FP8
(block-wise) vs Qwen3-8B BF16 baseline (perplexity 4.23 vs 4.33).
Loading custom-architecture checkpoints (Nemotron Omni family) against
transformers 5.14 surfaced API drift and wrapper mismatches:
- flash_attention_2 pinned in config but flash-attn not installed:
recursively fall back to eager attention, incl. nested sub-configs
(multimodal llm_config).
- all_tied_weights_keys shim: expected by the accelerate device-map
integration and assigned by 5.x post_init, but absent on older remote
code. Settable property aggregating legacy _tied_weights_keys,
filtering keys that don't resolve on multimodal wrappers.
- prepare_inputs_for_generation: 5.x may pass cache_position=None;
older remote code assumes a tensor. Wrap and synthesize it.
- Legacy list-style _tied_weights_keys normalized post-load so
save_pretrained works (emptied when tie_word_embeddings is false).
- Multimodal wrappers whose forward() requires media inputs are
unwrapped to their language_model submodule for text-only surgery.
- Register nemotron_h architecture (hybrid Mamba/attention/MoE, all
layer content under layer.mixer) in strategies/utils.py.
Raja Mukerji