feat: add KV-cache support for incremental decode#10
Open
GoooIce wants to merge 1 commit into
Open
Conversation
ShadowPEFT previously forced use_cache=False in all forward passes, requiring full-sequence recompute at every generation step. This was an implementation limitation, not a mathematical one — the injection and update modules are per-token pure functions with no cross-token recurrence. This commit enables cached generation: - ShadowPeftModel: add _shadow_cache internal state, conditionalize cache disabling in forward/forward_with_shadow, add shadow_cache parameter to _compute_initial_shadow_hidden - ShadowForCausalLM: remove forced use_cache=False, fix generation hooks (prepare_inputs_for_generation now slices input_ids for decode steps), propagate past_key_values in output, add get/set_experts_implementation for transformers 5.x compatibility - Add experiment/verify_cached_decode_alignment.py PoC (maxdiff ~1e-7) - Add tests/test_cached_decode.py (6 tests: generate alignment, prefill+decode, step-by-step logits, backward compat, shadow logits) Verified: generate(use_cache=True) produces identical token sequences to generate(use_cache=False). Default use_cache=False path is unchanged.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
ShadowPEFT previously forced use_cache=False in all forward passes, requiring full-sequence recompute at every generation step. This was an implementation limitation, not a mathematical one — the injection and update modules are per-token pure functions with no cross-token recurrence.
This commit enables cached generation:
Verified: generate(use_cache=True) produces identical token sequences to generate(use_cache=False). Default use_cache=False path is unchanged.