Add general MTP (multi-token-prediction head) pruning support - #28
Open
aquilarubra wants to merge 3 commits into
Open
Add general MTP (multi-token-prediction head) pruning support#28aquilarubra wants to merge 3 commits into
aquilarubra wants to merge 3 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>
MTP shares the main model's global expert count but has an independently learned router+expert pool. Adds mtp_util.py, which builds a real Qwen3_5MoeDecoderLayer for MTP's saliency pass rather than reimplementing attention by hand, and captures the real final-hidden-state/mask/position inputs the model's last block received (layerwise_observer.py/layerwise_prune.py) so that replay is faithful. Automatically a no-op when a checkpoint doesn't ship an MTP head. 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.
Summary
Depends on #26 (disk-streaming) — stacked on top of it, so the diff includes those commits until it merges; only the last commit is new here.
MTP heads share the main model's global expert count but have an independently learned router+expert pool, which the existing observer/pruning path doesn't touch at all.
What's in this PR
src/reap/mtp_util.py(new): builds a realQwen3_5MoeDecoderLayerfor MTP's saliency pass rather than reimplementing attention by hand.src/reap/layerwise_observer.py/src/reap/layerwise_prune.py: captures the real final-hidden-state/mask/position inputs the model's last block received, so MTP's saliency pass can replay them faithfully instead of recomputing RoPE/masks independently.Automatically a no-op when a checkpoint doesn't ship an MTP head.
Validation
Built a tiny synthetic checkpoint with real MTP tensors in the actual unfused per-expert-file layout and ran the full pipeline end-to-end: main-model observer → MTP replay capture → MTP's own saliency pass → combined disk-surgery pruning both pools consistently (same retained expert count, MTP experts renumbered sequentially with no gaps). Verified the "no MTP" case is a true no-op (falls through to the main-model-only path unchanged, no MTP-related file ever touched).