Add vision_only observer mode for text/vision expert-usage skew analysis - #31
Open
aquilarubra wants to merge 5 commits into
Open
Add vision_only observer mode for text/vision expert-usage skew analysis#31aquilarubra wants to merge 5 commits into
aquilarubra wants to merge 5 commits into
Conversation
Registers Qwen3_5MoeForConditionalGeneration in MODEL_ATTRS and generalizes get_moe() to walk model.model.language_model.layers, not just model.model.layers, for this architecture's nested multimodal wrapper. Adds a matching observer hook config (num_experts/top_k live on .experts/.gate submodules rather than the MoE block itself here). Generalizes the fused-expert pruning branch, which was hardcoded to Llama4's moe.router attribute and nn.Linear-style router with out_features, to also support Qwen3.5's .gate raw weight Parameter. Extracts select_retained_experts() as a standalone function from prune()'s inline body so the same selection logic can be reused outside an in-memory model (disk-based surgery). Signed-off-by: Fabrizio del Tin <devotedmystic@gmail.com>
REAP's layerwise-observer path was RAM-bounded only in appearance: the observer phase loaded the entire model onto CPU via device_map="cpu" before any block-wise logic ran, so peak RSS still scaled with total checkpoint size rather than one layer's size. This makes it possible to prune a checkpoint far larger than available RAM+VRAM. - New src/reap/disk_stream_util.py: reads/writes checkpoint tensors directly from safetensors shards, opening and closing each shard scoped to a single read (deliberately never caches file handles, since a held mmap handle keeps its touched pages resident in RSS). - layerwise_prune.py: replaces the full-RAM device_map="cpu" load with accelerate.init_empty_weights() (meta tensors) plus real materialization of only the embedding table, for local checkpoint directories (Hub model ids keep the original load path). - layerwise_observer.py: generalizes the fused_experts branch's router lookup (was hardcoded to moe.router) and adds a dense all-experts-all-tokens fallback for architectures whose experts need top-k routing args to call (Qwen3.5, unlike Llama4's simpler dense callable convention). Adds dual-mask support (ReplayBatch.causal_mask) so hybrid linear/full-attention models get the correct mask per layer type instead of replaying whichever mask block 0 received. Extends _move_block()/_offload_current_block() to materialize/free a block's real tensors from disk on demand via the new streaming primitive, instead of only supporting already-CPU-resident blocks. Signed-off-by: Fabrizio del Tin <devotedmystic@gmail.com>
transformers >=5.10 dispatches the fused-experts forward to torch._grouped_mm when available, whose meta registration hard-requires BF16 inputs. This layerwise replay runs in float32 (RoPE cos/sin promote hidden states through the chain), so force the eager per-expert loop instead, which is dtype-agnostic. Observer throughput is dominated by disk streaming, not the expert matmul dispatch, so this has no meaningful performance cost. Signed-off-by: Fabrizio del Tin <devotedmystic@gmail.com>
REAP's expert saliency is normally measured on text-only calibration, so experts that predominantly serve image tokens look unsalient and get pruned first even on VLM checkpoints. Adds an optional --vision_dataset (plus --vision_samples/--vision_split) LayerwiseArgs flag: when set, image+text calibration batches (from a local image directory or an HF dataset with an 'image' column) are appended to the text batches, and the vision tower is materialized from the disk- stream index so they can run. Off by default (text-only, matching existing behavior). New src/reap/vision_calib.py builds one processor batch dict per image sample; LayerwiseMoEObserver's existing input-forwarding already passes arbitrary tensor-valued dict entries into model(**batch), so no changes were needed to the observer/replay path itself. Signed-off-by: Fabrizio del Tin <devotedmystic@gmail.com>
Skips text calibration batches entirely when set (requires vision_dataset), so the observer state reflects only vision-token routing. Lets a new driver script diff a vision-only observer pass against a normal text-only pass to find which experts are vision-skewed, and at what keep_ratio REAP's pruning starts cutting into text-load-bearing experts instead of just vision ones. Signed-off-by: Fabrizio del Tin <devotedmystic@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Depends on #30 — stacked on top of it (uses the
vision_dataset/vision_samplesfields it adds).
What
Adds a
--vision_onlyflag toLayerwiseArgs(args.py) that, when set,skips text calibration batches entirely — the observer state reflects only
the
--vision_datasetsamples (requires it to be set).Why
Layerwise pruning's expert saliency is measured on whatever calibration data
you feed it. If you calibrate text-only (the default/upstream behavior),
experts that predominantly serve vision tokens look unsalient and get pruned
first — sometimes desirable (a text-only serving target), sometimes not
(you want to keep some multimodal capability).
Deciding how aggressively you can lean on that effect currently requires
guessing a
--compression_ratio/keep_ratioand hoping. With--vision_only,you can run the observer twice — once text-only, once vision-only — and diff
per-expert usage share between the two runs to see, empirically, which
experts are vision-skewed for this checkpoint's actual routing statistics,
before picking a pruning ratio. (We use this internally to build a
prune-ratio recommendation report; happy to share that tooling too if useful
upstream, but it's built around our own disk-streaming driver scripts so
didn't include it here to keep this PR minimal.)
Test plan
--vision_onlywithout--vision_datasetraises a clear errordata_batchesreflects only the vision samples (textbatch preparation is skipped)
--vision_only(defaultFalse), behavior is unchanged fromthe current
vision_dataset-only opt-in