Repository navigation

Capture resident KV aliases in recurrent CUDA Graphs - #96

Closed

0z5a wants to merge 81 commits into

ThinkFlowLab:mainfrom

0z5a:perf/alias-cuda-graph

Closed

Capture resident KV aliases in recurrent CUDA Graphs#96
0z5a wants to merge 81 commits into
ThinkFlowLab:mainfrom
0z5a:perf/alias-cuda-graph

Commits on Oct 2, 2026
Commits on Oct 4, 2026
- Add experimental LAST_EXITED alias storage and correctness checks
  0z5a
  committed
  63fab27
- Qualify official Ouro inputs and separate diagnostic and timed runs
  0z5a
  committed
  3763071
- Report three official Ouro pilot pairs and their measured regression
  0z5a
  committed
  9894e0e
- Add experimental record-granular KV allocation with conservative admission
  0z5a
  committed
  74ea87a
- Record CUDA compact allocation and shared-reader regression gates
  0z5a
  committed
  e2738be
- Validate compact records on the full Ouro checkpoint and record its live footprint
  0z5a
  committed
  1fdfe32
- docs: attribute alias pilot overhead with Nsight evidence
  0z5a
  committed
  51004fe
- perf: use direct reads before each depth first KV alias
  0z5a
  committed
  87504bb
- test: preserve EOS and qualified text stops for quality capture
  0z5a
  committed
  f068dd1
- test: pin loop bounds for full-depth quality controls
  0z5a
  committed
  f37bf01
- docs: retain compact official E2E pilot regression
  0z5a
  committed
  41816d0
- Merge branch 'perf/compact-kv-records' into perf/hybrid-kv-reader
  0z5a
  committed
  567ba5d
- test: coordinate checkpoint IO independently of quality execution
  0z5a
  committed
  32e9c02
- docs: record hybrid reader correctness and pending E2E validation
  0z5a
  committed
  6224926
- perf: batch compact KV descriptors into two host transfers
  0z5a
  committed
  3d9f40f
- fix: avoid reserving fixed compact prefill capacity twice
  0z5a
  committed
  deb21b2
- Merge branch 'perf/compact-kv-records' into perf/compact-kv-staging
  0z5a
  committed
  a5e7419
- feat: reclaim skipped past KV credits while reserving every future depth
  0z5a
  committed
  861f799
- Merge branch 'perf/compact-kv-records' into perf/hybrid-kv-reader
  0z5a
  committed
  7ca67d6
- test: complete credit-limited engine workloads through full-depth arrivals
  0z5a
  committed
  dd37dd8
- docs: retain GSM8K100 failure and exact corrected compact parity
  0z5a
  committed
  4e497fb
- docs: compact raw quality artifacts without changing frozen inputs
  0z5a
  committed
  c1a3d76
- Merge branch 'perf/compact-kv-records' into perf/hybrid-kv-reader
  0z5a
  committed
  756614d
- docs: link corrected compact quality evidence
  0z5a
  committed
  cdc61b6
- docs: record exact official hybrid quality comparison
  0z5a
  committed
  24a4bed
- Merge branch 'perf/compact-kv-records' into perf/compact-kv-staging
  0z5a
  committed
  7369b81
- docs: record compact descriptor CPU and CUDA validation
  0z5a
  committed
  a2361f2
- Merge branch 'perf/compact-kv-staging' into perf/reclaim-skipped-credits
  0z5a
  committed
  93ede6e
- test: add finite CUDA gate for reclaimed credits under full-depth pressure
  0z5a
  committed
  cf70584
- Document confirmed-skip admission bounds and pressure gates
  0z5a
  committed
  9b252b1
- Report all five hybrid reader E2E pilot shapes
  0z5a
  committed
  2e3dac4
- Record CUDA admission pressure and full-depth completion
  0z5a
  committed
  1fab7f8
- Record official fixed-memory admission and exit differences
  0z5a
  committed
  2118175
- Merge branch 'perf/hybrid-kv-reader' into test/looped-model-coverage
  0z5a
  committed
  0f9c7db
- Qualify standalone Nanbeige checkpoint loading
  0z5a
  committed
  1e355ce
- Integrate existing Huginn runtime with boundary-aware alias lifetime
  0z5a
  committed
  30fcaf4
- Qualify tied Huginn weights and persistent RoPE buffers
  0z5a
  committed
  6809cbb
- Record Thinking batch coverage and native model qualification
  0z5a
  committed
  300910f
- Add finite Nanbeige and Huginn cache mode gate
  0z5a
  committed
  5d49b2a
- Record official checkpoint parity for packed descriptors
  0z5a
  committed
  2d21051
- Merge branch 'perf/compact-kv-staging' into test/looped-model-coverage
  0z5a
  committed
  777e255
- Publish completed alias depths once per synchronous batch
  0z5a
  committed
  0ec6c69
- Document grouped publication checks and pending CUDA validation
  0z5a
  committed
  5ef8786
- Route rectangular aliases through resident async banks
  0z5a
  committed
  5023ef4
- Order resident alias readers after allocation reset
  0z5a
  committed
  b5358b9
- Add cross-stream alias lifetime qualification probes
  0z5a
  committed
  2cd0a7b
- Use a full-depth prefill entry in the lifetime probe
  0z5a
  committed
  6ec1e85
- Qualify batched exit publication across CUDA tile boundaries
  0z5a
  committed
  c0180b7
- Merge branch 'perf/batched-kv-finalization' into perf/async-alias-routing
  0z5a
  committed
  28fd635
- Count actual batch sizes and submitted loop work in timed captures
  0z5a
  committed
  6d21353
- Capture resident rectangular alias attention in CUDA Graphs
  0z5a
  committed
  4adada8
- Add single-load official delayed and Graph comparison driver
  0z5a
  committed
  80a9972
- Record Nanbeige and Huginn CUDA cache qualification
  0z5a
  committed
  40bd7ec
- Merge branch 'test/looped-model-coverage' into perf/batched-kv-finalization
  0z5a
  committed
  dc36cc3
- Record CUDA grouped publication and cache regression results
  0z5a
  committed
  67dd7a4
- Merge branch 'perf/batched-kv-finalization' into perf/async-alias-routing
  0z5a
  committed
  72b08e6
- Record CUDA async alias ordering and reuse qualification
  0z5a
  committed
  9389dd1
- Merge branch 'perf/async-alias-routing' into perf/alias-cuda-graph
  0z5a
  committed
  01be75a
- Record actual alias CUDA Graph captures and replay parity
  0z5a
  committed
  cf1f58a
- Include compact ownership in single-load model diagnostics
  0z5a
  committed
  f2b6add
Commits on Oct 5, 2026

Commits

Commits on Oct 2, 2026

Commits on Oct 4, 2026

Commits on Oct 5, 2026