Skip to content

CHANGELOG.md

Latest commit

 

History

History
204 lines (175 loc) · 9.82 KB

CHANGELOG.md

File metadata and controls

204 lines (175 loc) · 9.82 KB
Raw

Changelog

All notable UI-surface changes since the V7 honest-closure pass.

V7-D-block: TileLang→Metal kernel epic closure (2026-05-23 14:30 GMT+2)

D-block (TileLang→Metal kernels, 8 subtickets + Path B adapter + epic

  • breakdown task = 13 tickets) closed as stale bd vs reality.
  • All 10 kernels shipped in cppmega_mlx/nn/_tilelang/: topk_selector (Path B + Path C, 923 LOC), sparse_mla / sparse_mla_path_c (BF16 fwd+bwd), sparse_mla_fp8 / sparse_mla_fp8_path_c (FP8 fwd+bwd), sparse_mla_blockscaled / sparse_mla_blockscaled_path_c (e8m0 fwd+bwd), mamba3 / mamba3_path_c (fwd+bwd via Path C), _mamba3_helpers_tilelang.compute_dacs_segsum (Triton-free).
  • Path B adapter (0ce): TileLangMetalAdapter + wrap_tilelang_metal_kernel in cppmega_mlx/nn/_tilelang/_mlx_runtime.py:628-876. Takes TileLang artifact → extracts MSL → mx.fast.metal_kernel.
  • Test fix: tests/test_path_c_sparse_mla_apply_emit_indent.py was calling _mamba3_fp8_train_acceptance_region(model_config=…) but the signature was simplified to include_backward only. Dropped the stale kwarg; 3 tests now PASS.
  • Regression evidence: 407/407 kernel tests PASS across 12 test files at commit 77a1404. bd open count: 57 → 44.

Closed tickets: cppmega-mlx-za1, za1.1, za1.2, za1.3, za1.4, za1.5, za1.6, za1.7, za1.8, za1.9, za1.10, za1.11, 0ce.

V7-M0.7 + M0.3 closure (2026-05-23 06:00 GMT+2)

  • M0.7 (resumable training) ✅ green — stage_train now auto-derives a sidecar opt-state file <ckpt>.opt when checkpoint_save_path is given without an explicit opt_state_save_path; symmetric auto-resolve on load. The sidecar bundles AdamW m/v moments + rng_key so a single "checkpoint" path yields full resumable semantics. Flipped tests/v4/test_m07_resumable_roundtrip.py from xfail-strict to PASS — resumed loss continues the baseline within <0.1 per step (was Δ≈0.75).
  • M0.3 (capture_logits wiring) ✅ MLX-side green — dry_forward(graph, ..., capture_logits=True, seed=N) surfaces output_logits (shape) + output_values (flat list) in extras. bench/m03_cuda_logits_parity_harness.py now writes bench/baselines/m03_mlx_logits.npy (shape [1, S, H]). Status: awaiting_cuda_ref — GB10 reference is the remaining external blocker (bd cppmega-mlx-uwhj).
  • M0.5 (FastMTP parity): refreshed bench/baselines/m05_mlx_fastmtp.json (mlx_loss=5.7456, weight_delta_norm=0.4046); still awaiting GB10 CUDA reference (bd cppmega-mlx-hjfn).
  • New regression: tests/v4/test_m03_capture_logits.py (4 tests) pins shape, finite-values, and same-seed determinism for the capture path.

V7-P-block: full-scale UI + convergence e2e + M0 baseline (2026-05-23)

  • P5: DimEnvEditor scale-preset dropdown — mini / dev_128 / small_512 / medium_1k / large_2k / llama3_8b / llama3_70b. Each preset self-consistent (nh*head_dim == H). Unblocks full-scale H=4096 through UI. testid dim-env-preset + per-option ids.
  • P6/P7: p6_real_convergence_curve.spec.ts asserts visible LossChart shows monotonic decline across 16 steps via median-quartile comparison. Replaces tautological "num_steps == N" checks.
  • P10: p10_hybrid_muon_adamw.spec.ts covers muon_adamw_hybrid optim end-to-end at H=512 via the P5 preset selector.
  • M0.6: bench/m06_memory_bench.py + generated bench/baselines/m06_memory.json — the missing artefact from bd t8f.6. Walks verify_and_estimate to produce per-brick params/activations bytes breakdown. CLI knobs --H --S --B --preset --depth --out so dev-128 grad-checkpoint+AdamW can re-run.
  • M0.7: initial honest finding (xfail-strict, Δ=0.75 loss divergence on resume). Superseded by the M0.7 closure entry above (sidecar opt-state landed 2026-05-23 06:00 GMT+2).
  • Deferred (need GB10/hardware/multi-hour): P8 multi-epoch UI e2e, P9 1k+ step Playwright run, M0.3 GB10 CUDA-side reference, M0.5 GB10 FastMTP CUDA reference. bd hjfn / uwhj / A05 carry the remaining external work; MLX side captured for both M0.3 + M0.5.

V7-N-block: deferred-work bd tickets + gen.run UI (2026-05-23)

  • gen.run UI: new Inference tab with GenerationPanel — prompt_tokens parser, sampler strategy selector (greedy / temperature / top_k / top_p), per-strategy hyperparams, smoke toggle, token chips + finish_reason + elapsed_ms surface. testid contract: gen-prompt-tokens / gen-strategy / gen-max-new-tokens / gen-temperature / gen-top-k / gen-top-p / gen-seed / gen-vocab-size / gen-smoke / gen-run / gen-result / gen-token-{i} / gen-finish-reason / gen-elapsed-ms / gen-error.
  • N01..N07 bd tickets created for xfail markers + TODO comments
    • TileLang-Metal kernel breakdown: gemm_softmax emitter wiring, sparse_mla path-C custom-op, triton OP_TABLE coverage (RFC §5.5), mamba3 wave-6 DSL port, dsa_splitk_indexer KL term, training._quantize_8bit dDequantizeBlockwise, za1 epic per-kernel breakdown.

V7-C-block: checkpoint robustness (closed via dedup, 2026-05-23)

C01..C06 had two parallel ticket trees — the closed cousins (q3jq/z6dj/91h1/9t1m/fazl/ylyh) already shipped the backend + TopBar ckpt-inspect-block + train_checkpoint_save/load paths + K7 RunHistoryPicker warm-start UX. The 6 open dupes (1q0, 9vv, 4whl, 8t3b, bjiy, 7n4u) closed with DEDUP notes pointing to the shipped twin.

V7-H-block tail: pause rehydrate + project run-id + redo defense + tokenizer roundtrip (2026-05-23)

  • H46: TokenizerPlayground byte-roundtrip pill from existing EncodeVisualizeResult.capabilities.byte_roundtrip bool. Green/red data-roundtrip=ok|fail testids.
  • H41: App.tsx useEffect on mount reads localStorage activeTrainRunId, hits pipeline.status, restores trainRunId + trainInFlight + trainPaused when running && !aborted. Stale ids scrubbed.
  • H42: trainRunId persisted to vbgui_active_train_run_id_<project-uuid> so concurrent project trains don't clobber.
  • H43: useHistory.redo() now returns {snapshot, rejected}; new markRejected() API tags current top. App.tsx runVerify flags any error-severity-gotcha snapshot; handleRedo surfaces a 4s toast when redo lands on a rejected one.

V7-L-block tail: GotchaPayload extensions (2026-05-23)

  • L48: GotchaPayload + GotchaState gain suggested_fix. verify RPC post-processes via backend _SUGGESTED_FIX_BY_ID lookup so new gotcha ids ship with their one-click fix label without a UI release. GotchasTab accepts EITHER backend suggested_fix OR legacy AUTO_FIXABLE id; when onAutoFix isn't passed it renders an italic hint.
  • L49: parseSourceFile() extracts the last path segment from GotchaState.reference; renders as a small monospace src: chip with the full path on hover.
  • L50: per-severity BG_TINT (red/amber/blue pastel) + a colored pill (data-severity attr) so warnings stop blending into info gotchas.

V7-M-block: train-extras UI surfacing (2026-05-23)

Backend was emitting 16 train-stage extras the UI silently buried in its generic JSON dl dump. Each got a dedicated visual surface.

  • M21 / M23: LossChart now overlays losses_smoothed (green) and val_losses (purple) on top of the primary losses curve. testids: extras-loss-chart-line / -line-smoothed / -line-val.
  • M22: dedicated LR chart (extras-lr-chart-svg, extras-lr-chart-line-lr) rendered from lr_trajectory.
  • M24: perplexity + bits_per_byte scalar badges.
  • M25: master_dtype + dtype_actual badges (shown side-by-side so request vs reality is one glance).
  • M26: fp8_active ON badge (suppressed when false).
  • M27: sharding-panel with sharding_applied + per-rank param bytes formatted via Number#toLocaleString.
  • M28: FIM badge with fim_ratio percent.
  • M29: side-channels-panel listing side_channels_observed with per-channel testid.
  • M30: per-brick grad-norm horizontal bars with data-grad-norm attribute for e2e assertion.
  • M31: brick_kinds pill row (handles array or comma-string).
  • M32: MoE dashboard — 9 keys (routing_entropy, load_balance_loss, per_expert_load bars, dropped/rerouted/ overflow ratios, capacity_per_expert, capacity_factor, num_experts).
  • M33: grad-clip activity panel (max_grad_norm_seen + clip count).
  • M34: optimizer_kind badge for the current run.
  • M35: gradient_reduce_ms badge (fake_ranks proxy).
  • M36: inference-steps flow trace in DimensionsTab (follow-up).

All surfaces share the existing HelpIcon pattern — clicking the "?" opens a modal with what/why/example.

V7-L-block: error/status visibility (2026-05-23)

  • L46: ErrorDetailsPanel parses RpcError.data.errors[] (Pydantic shape) + traceback + stage metadata, instead of dropping the data blob and showing only error.message.
  • L47: RunResultModal pre-expands the first fail stage on render and adds a red outline + light-red row background to the failing row (data-first-failed='true').

V7-K-block: 10 missing train controls (2026-05-23)

TrainOptionsPanel (K3-K6, K8) + RunHistoryPicker (K7) + TrainLiveControls (K9, K10) + dim_env-into-train wiring (K2) + the already-present TopBar num_steps input (K1). 7 Playwright e2e green. See vbgui/e2e/scenarios/k*.spec.ts.

V7-F-block: brick constructor (2026-05-23)

LossChart foundation + DimEnvEditor with snap-to-valid fixes (F56b) + GalleryTab sortable cache (F58) + SweepPanel H-sweep overlay (F53) + TokenizerMatrixTab (F55) + BrickContextPanel swap-kind (F52) + TransplantBar (F54) + InsertIntoEdgeBar (F51)

  • ParallelComposeBar (F57). 8 Playwright e2e green.

V7-J general-purpose surfacing primitives

  • HelpIcon + HELP_TOPICS — central registry of explanation modals; every numeric / structural surface that lands in this changelog ships with a ? icon linked to a topic entry.