Repository navigation

Research release: SDPA, KV-cache experiments and reproducible evaluation - #1

Merged

AnvitDevadiga merged 47 commits into

mainfrom

experiments/sdpa-training-v2

Oct 22, 2026

Merged

Research release: SDPA, KV-cache experiments and reproducible evaluation#1
AnvitDevadiga merged 47 commits into
mainfrom
experiments/sdpa-training-v2

Commits on Oct 22, 2026
- Add selectable attention backends, fix half-precision RoPE, add correctness tests
  
  AnvitDevadiga
  and
  claude
  committed
  274722f
- Add attention benchmark, quality comparison, provenance records and report builder
  
  AnvitDevadiga
  and
  claude
  committed
  615767d
- Measure MPS peak memory per op; fix MPS float64 cast; add training records
  
  AnvitDevadiga
  and
  claude
  committed
  0502118
- Record attention, quality and training runs; publish results and audit
  
  AnvitDevadiga
  and
  claude
  committed
  274fd83
- Add provenance tests; document attention study and limitations in README
  
  AnvitDevadiga
  and
  claude
  committed
  cc1c822
- Add cached vs uncached decoding benchmark with window and memory scenarios
  
  AnvitDevadiga
  and
  claude
  committed
  93507c4
- Record cached vs uncached decoding on CPU fp32, MPS fp32 and MPS fp16
  
  AnvitDevadiga
  and
  claude
  committed
  d471a7d
- Fix float64 downcast in reference attention; add gradient and dropout checks
  
  AnvitDevadiga
  and
  claude
  committed
  e6cc478
- Add selectable attention backend to training with timing and memory capture
  
  AnvitDevadiga
  and
  claude
  committed
  29a5c15
- Add matched SDPA vs default attention training experiment script (v1 corpus)
  
  AnvitDevadiga
  and
  claude
  committed
  48e95f8
- Record matched default-vs-SDPA training on science-v1 (3 seeds) and corrected-oracle quality
  
  AnvitDevadiga
  and
  claude
  committed
  a0d9db8
- Add matched training-step benchmark; report KV cache and training sections
  
  AnvitDevadiga
  and
  claude
  committed
  d3e6f26
- Record matched training-step benchmark on CPU and MPS (step 1 frozen)
  
  AnvitDevadiga
  and
  claude
  committed
  827b605
- Add corpus protocol science-v2: deterministic cleaning, splits, untouched test book
  
  AnvitDevadiga
  and
  claude
  committed
  a828f2d
- Add science-v2 corpus files, integrity/contamination tests and protocol script
  
  AnvitDevadiga
  and
  claude
  committed
  897db37
- Record science-v2 training (3 seeds x 2 backends) and one-time test evaluation
  
  AnvitDevadiga
  and
  claude
  committed
  2f23f69
- Report SDPA decision evidence and science-v2 results separately from v1
  
  AnvitDevadiga
  and
  claude
  committed
  23f443b
- Make SDPA the default attention backend; name the legacy path mha
  
  AnvitDevadiga
  and
  claude
  committed
  2dc1f73
- Reproducibility: v1 corpus regeneration, dependency lock, record validation
  
  AnvitDevadiga
  and
  claude
  committed
  4b6803f
- Let legacy KV-cache and INT8 scripts write validated experiment records
  
  AnvitDevadiga
  and
  claude
  committed
  13530f4
- Record legacy v1 KV-cache and INT8 re-measurements (mha backend, CPU)
  
  AnvitDevadiga
  and
  claude
  committed
  52441b4
- Release documentation: README restructure, SDPA default, test-set status
  
  AnvitDevadiga
  and
  claude
  committed
  29e8574
- CI: full-history checkout and locked dependencies; store repo-relative paths
  
  AnvitDevadiga
  and
  claude
  committed
  614977b
- Add release checklist with passing checks, corrections and outstanding items
  
  AnvitDevadiga
  and
  claude
  committed
  c256f20
- Finding 1: deterministic model identity digest separate from checkpoint file hash
  
  AnvitDevadiga
  and
  claude
  committed
  cd6a11d
- Findings 2 and 3: experiment identity, supersession rules, strict evaluation provenance
  
  AnvitDevadiga
  and
  claude
  committed
  3a945bd
- Finding 4: persist raw per-repeat KV-cache timings and derive summaries from them
  
  AnvitDevadiga
  and
  claude
  committed
  57be26d
- Finding 4: replacement KV-cache records with raw per-repeat samples
  
  AnvitDevadiga
  and
  claude
  committed
  753ff59
- Regression tests for findings 1-4; resolve same-identity replicate checkpoints
  
  AnvitDevadiga
  and
  claude
  committed
  8a04f91
- Correction B: device-correct peak-memory measurement; never host RSS for CUDA
  
  AnvitDevadiga
  and
  claude
  committed
  1c33637
- Finding 1: attest model identities for training records that predate them
  
  AnvitDevadiga
  and
  claude
  committed
  935a229
- Record model-identity attestations for 24 historical checkpoints
  
  AnvitDevadiga
  and
  claude
  committed
  a9780e3
- Corrections A and D: qualify the float64 oracle; release verification script
  
  AnvitDevadiga
  and
  claude
  committed
  8740092
- Docs: record schema and provenance rules; reconcile audit corrections
  
  AnvitDevadiga
  and
  claude
  committed
  c73ea2d
- Tests: reject tampered identity attestations and enforce attested identities
  
  AnvitDevadiga
  and
  claude
  committed
  d456295
- Release checklist: independent audit findings, validation evidence, number changes
  
  AnvitDevadiga
  and
  claude
  committed
  ad71757
- Fix release blockers 1-3 in record validation
  
  AnvitDevadiga
  and
  claude
  committed
  a8e300d
- Report renders derived KV speedups and requires strict identity evidence
  
  AnvitDevadiga
  and
  claude
  committed
  58b0782
- Regression tests for the three release blockers; validator artifact/baseline options
  
  AnvitDevadiga
  and
  claude
  committed
  439a003
- Docs: identity resolution, evidence levels, baseline trust assumptions, derived KV ratios
  
  AnvitDevadiga
  and
  claude
  committed
  0c84b30
- Release checklist: release-blocker fixes, probe results, identity evidence and trust
  
  AnvitDevadiga
  and
  claude
  committed
  6748924
- Docs: minimal README with monochrome light/dark architecture diagrams
  
  AnvitDevadiga
  and
  claude
  committed
  02a902f
- Diagram generator: rename shadowed loop variable (ruff PLR1704); output unchanged
  
  AnvitDevadiga
  and
  claude
  committed
  160aa65
- CI: gate MPS tests on a real capability probe; run full release verification
  
  AnvitDevadiga
  and
  claude
  committed
  3af16f6
- Merge experiments/sdpa-training-v2 (CI MPS capability gating) into docs/readme-diagrams
  
  AnvitDevadiga
  and
  claude
  committed
  533e7c9
- README: select the architecture diagram by GitHub theme, not OS theme
  
  AnvitDevadiga
  and
  claude
  committed
  5da89d7
- Merge pull request #2 from AnvitDevadiga/docs/readme-diagrams
  AnvitDevadiga
  authored
  cd90331

Commits

Commits on Oct 22, 2026