fix(ruvLLM): P4.1 wire trainer backprop endpoint gradients#416
Open
Stricttype wants to merge 5 commits intoruvnet:mainfrom
Open
fix(ruvLLM): P4.1 wire trainer backprop endpoint gradients#416Stricttype wants to merge 5 commits intoruvnet:mainfrom
Stricttype wants to merge 5 commits intoruvnet:mainfrom
Conversation
- examples/ruvLLM/Cargo.lock - examples/ruvLLM/Cargo.toml - examples/ruvLLM/docs/api-reference.md - examples/ruvLLM/docs/code-standards.md - examples/ruvLLM/docs/codebase-summary.md - examples/ruvLLM/docs/configuration-guide.md - examples/ruvLLM/docs/deployment-guide.md - examples/ruvLLM/docs/handoffs/2026-05-02-1943-auto.md - examples/ruvLLM/docs/project-overview-pdr.md - examples/ruvLLM/docs/system-architecture.md - examples/ruvLLM/docs/testing-guide.md - examples/ruvLLM/learn/260502-1900-init-ruvllm/learn-results.tsv - examples/ruvLLM/learn/260502-1900-init-ruvllm/summary.md Co-Authored-By: Pi Coding Agent <pi@localhost>
- examples/ruvLLM/Cargo.lock - examples/ruvLLM/config/example.toml - examples/ruvLLM/config/pretrain.toml - examples/ruvLLM/scripts/fetch-simple-wiki.sh - examples/ruvLLM/src/bin/pretrain.rs - examples/ruvLLM/src/bin/sidecar.rs - examples/ruvLLM/src/config.rs - examples/ruvLLM/src/lib.rs - examples/ruvLLM/src/sona/mod.rs - examples/ruvLLM/src/sona/persist.rs - examples/ruvLLM/src/training.rs - examples/ruvLLM/tests/persist_integration.rs Co-Authored-By: Pi Coding Agent <pi@localhost>
- rename src/data/ → src/corpus/ (gitignore conflict: data/ pattern blocks Rust source) - add corpus module: wiki corpus iter, tokenizer wrapper, tokenized dataset - add tests/wiki_pipeline_test.rs (5/5 PASS) - surgical fixes for pre-existing candle 0.8 API drift in src/inference_real.rs - add From<candle_core::Error> shim in src/error.rs (unblocks --features real-inference) - extend src/training.rs: DatasetSource trait, ModelCheckpoint serde, save_checkpoint, measure_baseline_perplexity P4 status: DONE_WITH_CONCERNS — pre-existing issues surfaced: - Trainer::train computes loss but does not call optimizer.step (no backprop) → perplexity-delta is structurally 0% until follow-up patch - SmallTransformer lacks from_checkpoint() constructor → trained checkpoints are saved but to_q4_weights() re-randomizes; follow-up needed to load saved checkpoint into inference path - TokenizerWrapper::from_pretrained stubbed (requires tokenizers/http feature, not currently enabled); inline whitespace WordLevel fallback works for offline pilot Smoke tests: - cargo check: PASS - cargo check --features persistence: PASS - cargo check --features real-inference: PASS - cargo check --features persistence,real-inference: PASS - cargo test --features persistence --test persist_integration: 4/4 PASS - cargo test --features real-inference --test wiki_pipeline_test: 5/5 PASS
Co-Authored-By: Pi Coding Agent <pi@localhost>
Trainer::train_epoch now computes analytical gradients for the output endpoint (cross-entropy → lm_head → RMSNorm) and applies optimizer.step to lm_head, output_norm, and embeddings. Transformer body remains a fixed feature extractor (endpoint-only approximation, Option A). - TrainableModel::forward_with_cache exposes the LM-head input (normed) required for analytical gradients without changing forward()'s signature. - Per-batch gradient accumulation with averaging + global L2 clipping. - TrainingMetrics.grad_norm now populated from accumulated batch norms. - Public Trainer API (new/train/train_epoch/into_model/metrics_history/ save_checkpoint_periodic/model) unchanged. Tests: - src/training.rs::tests::test_train_epoch_decreases_loss — ≥10 steps on synthetic data, asserts final < initial loss. - src/training.rs::tests::test_train_epoch_updates_lm_head — verifies optimizer actually mutates lm_head between epochs. - tests/wiki_pipeline_test.rs::test_perplexity_5pct_floor_with_backprop — fixture corpus, ≥1 epoch, asserts final_ppl < 0.95 * initial_ppl. Observed: 51.74 → 45.50 (ratio 0.880, 12% improvement). - Existing test_perplexity_better_than_random tightened from non- regression (≤2.0) to improvement (<0.95). cargo check (default + --features real-inference): pass cargo test --lib --features real-inference: 105 passed cargo test --test wiki_pipeline_test --features real-inference: 6 passed
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
Follow-up to #414. Fixes the structural blocker on Acceptance #4 ("perplexity better than random-init baseline") by wiring actual gradient updates into
Trainer::train_epoch.Bug:
train_epochcomputedcompute_lossbut never calledoptimizer.stepagainst any weight matrices — model parameters never changed regardless of training duration.Approach: endpoint-only backprop (Option A)
Hand-rolled analytical gradients flow through cross-entropy →
lm_head→ final RMSNorm, with a coarse identity-bypass for embeddings. Body (attention QKV, FFN, per-layer norms) remains random-init, treated as a fixed feature extractor.~150 LOC, no new deps. Public API unchanged:
Trainer::{new, train, train_epoch, into_model, metrics_history, save_checkpoint_periodic, model}andTrainableModel::{save/load_checkpoint, from/to_checkpoint, to_q4, forward, compute_loss}keep their signatures.forward_with_cacheis purely additive.Acceptance evidence
cargo check(default)cargo check --features real-inferencecargo test --lib --features real-inferencecargo test --test wiki_pipeline_test --features real-inferencetest_train_epoch_decreases_loss(≥10 steps, synthetic)test_train_epoch_updates_lm_head(mutation sanity)test_perplexity_5pct_floor_with_backpropPerplexity (fixture, 2 epochs, lr=1e-2): 51.736 → 45.504 (ratio 0.880, 12% reduction — well past 5% floor).
Caveats / future work
wq/wk/wv/wo/w1/w2/w3+ per-layer norms) intentionally remain at random-init. Endpoint-only by design.output_normgradient recovershidden/rmsfromnormed/g; defensive guard handles the rare zero case.Test plan
--features real-inference--features persistence(orthogonal patch)Depends on / built on top of #414.
🤖 Generated with Claude Code