Repository navigation

Stream AutoScheme's per-layer sensitivity scoring block-by-block - #2063

Merged

xin3he merged 28 commits into

intel:mainfrom

aquilarubra:pr/autoscheme-streaming

Jul 30, 2026

Merged

Stream AutoScheme's per-layer sensitivity scoring block-by-block#2063
xin3he merged 28 commits into
intel:mainfrom
aquilarubra:pr/autoscheme-streaming

Commits on Jul 29, 2026
- Add AR_DISK_STREAM_MODEL and AR_RESUME_DIR env vars
  aquilarubra
  committed
  dac780c
- Add disk-streaming primitive for per-block checkpoint materialization
  aquilarubra
  committed
  083bb44
- context/model: build meta-device skeleton under AR_DISK_STREAM_MODEL
  aquilarubra
  committed
  eaaf80c
- offload: materialize/reload blocks from meta correctly
  aquilarubra
  committed
  c90acd4
- offload: use a deterministic offload dir when AR_RESUME_DIR is set
  aquilarubra
  committed
  a72847b
- compressors/base: propagate model_dir to offloader; defer resume clear
  aquilarubra
  committed
  011874a
- compressors/data_driven: honor disk streaming outside resumability
  aquilarubra
  committed
  97696fe
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  
  pre-commit-ci[bot]
  authored and
  aquilarubra
  committed
  a863e84
- Strip leaked internal-repo comments; fix meta-materialize dtype bug
  aquilarubra
  committed
  933cb6b
- Add unit tests for disk_stream_util.py primitives
  aquilarubra
  committed
  09450c8
- Re-trigger CI (unrelated HF Hub 503 flake on test_gptoss)
  aquilarubra
  committed
  de38945
- Re-trigger CI (retest after flaky VRAM/RAM perf checks and llmc_format flake)
  
  Fabrizio del Tin
  authored and
  aquilarubra
  committed
  b60e864
- Re-trigger CI (unrelated HF Hub 503 flake on test_gptoss)
  aquilarubra
  committed
  22e0c2d
- Re-trigger CI (retest after flaky VRAM/RAM perf checks and llmc_format flake)
  
  Fabrizio del Tin
  authored and
  aquilarubra
  committed
  06f1a0b
- Add AR_DISK_STREAM_MODEL and AR_RESUME_DIR env vars
  aquilarubra
  committed
  949ff2c
- Re-trigger CI (unrelated HF Hub 503 flake on test_gptoss)
  aquilarubra
  committed
  0050b4d
- compressors/data_driven: resumable standard tuning loop (AR_RESUME_DIR)
  aquilarubra
  committed
  11d1d45
- Add ResumeState: crash/resume checkpointing for the tuning loop
  aquilarubra
  committed
  d9cc0fc
- shard_writer: discover pre-existing shards to resume across processes
  aquilarubra
  committed
  f595ce1
- export: skip pack_layer() for a resumed, still-meta layer
  aquilarubra
  committed
  3140a5f
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  
  pre-commit-ci[bot]
  authored and
  aquilarubra
  committed
  e0ca142
- Strip leaked internal-repo comment markers
  aquilarubra
  committed
  d7f0e36
- Add unit and integration tests for AR_RESUME_DIR
  aquilarubra
  committed
  995afea
- Re-trigger CI (retest after flaky VRAM/RAM perf checks and llmc_format flake)
  
  Fabrizio del Tin
  authored and
  aquilarubra
  committed
  ead1eda
- Fix envs.py syntax corrupted during rebase conflict resolution
  aquilarubra
  committed
  2092f07
- auto_scheme: disk-streaming support prep (gen_auto_scheme low_cpu_mem_usage, docs, tests)
  aquilarubra
  committed
  7434f22
- auto_scheme: stream per-layer sensitivity scoring block-by-block
  aquilarubra
  committed
  b9eb98f
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  pre-commit-ci[bot]
  committed
  8f14123

Commits

Commits on Jul 29, 2026