This project publishes no release notes. Between v0.5.3 and v0.5.4 there were 4 commits, 3 of them substantive:

• Detect Windows model storage speed
• Simplify Windows storage detection
• Simplify Windows disk classification

This project publishes no release notes. Between v0.5.2 and v0.5.3 there were 15 commits, 10 of them substantive:

• hip: prefer the ROCm 7 runtime and correct dispatch symbol resolution
• shmem-detect: drop vendor specific sysmem fallback wording
• control: prefer torch.version.hip over the local version segment
• vrambuf: lock and tear down the rocm windows VA reservation pool
• control: budget integrated rocm GPUs against system RAM on linux
• Add runtime simple_vram_headroom accessors
• Refuse a second init_devices() before plat_init()
• Test simple_vram_headroom changes at runtime
• Allow non-graph frees during memory compilation
• Report malloc graph rogue count

This project publishes no release notes. Between v0.5.1 and v0.5.2 there were 2 commits, 1 of them substantive:

• Try to fix hard crash on workflow cancel.

This project publishes no release notes. Between v0.5.0 and v0.5.1 there were 2 commits, 1 of them substantive:

• Abort incomplete malloc graphs safely

This project publishes no release notes. Between v0.4.15 and v0.5.0 there were 42 commits, 41 of them substantive:

• Add native CUDA malloc graph compiler
• Expose malloc graph Python API
• Test malloc graph allocation and nesting
• Test malloc graph replay contract
• Test malloc graph ownership errors
• Allow VBAR eviction during CUDA capture
• Evict VBARs for malloc graph pages
• Compile small CUDA allocations
• Add malloc graph iteration API
• Pause malloc graph recording
• Split physical page state
• Use scope depth for allocation ownership
• Separate virtual page mappings
• Snapshot allocation state at scope entry
• Link allocation events backwards
• Rebuild allocation state from scope snapshots
• Store event successors in arrays
• Branch malloc graph replay on cache misses
• Unify malloc graph entry under push
• Report malloc graph breaks

…and 21 more.

This project publishes no release notes. Between v0.4.14 and v0.4.15 there were 2 commits, 1 of them substantive:

• Set Windows NVML pressure headroom to 512 MB

This project publishes no release notes. Between v0.4.13 and v0.4.14 there were 6 commits, 3 of them substantive:

• docs: note AMD/ROCm support in README
• fix: fail CLA allowlist step on API errors
• Set Windows NVML pressure headroom to 768 MB

This project publishes no release notes. Between v0.4.12 and v0.4.13 there were 2 commits, 1 of them substantive:

• Support GPU copies on the default stream

This project publishes no release notes. Between v0.4.11 and v0.4.12 there were 9 commits, 7 of them substantive:

• remove pull request template
• add CHECK_CU_OOM_ERROR
• Add ERROR logging for all possible failure paths of read_file_slice
• Add ERROR logging for all possible failure paths of read_file_slice
• Add CHECK_CU_ERROR logging
• Route native logs through Python logging
• fix namespace collision

This project publishes no release notes. Between v0.4.10 and v0.4.11 there were 15 commits, 9 of them substantive:

• ci: add CLA Assistant workflow
• recycle VRAMBuffer VA reservation to avoid HIP VMM crash
• Make VA pool lock initialization thread-safe on ROCm/Windows
• Move VA pool to per-device context on ROCm/Windows
• Enhance CLA workflow with author-only allowlist
• Fix commit author login extraction in CLA workflow
• Remove sequential scan hint
• prioritize: add optional clamp on async pre-emptive free
• prioritize NVML pressure

This project publishes no release notes. Between v0.4.9 and v0.4.10 there were 4 commits, 3 of them substantive:

• model-vbar: double-sync
• control: add extra_vram_headroom
• implement simple_vram_headroom setting

This project publishes no release notes. Between v0.4.8 and v0.4.9 there were 4 commits, 3 of them substantive:

• set watermark
• allocations dirty reporting
• vbar_fault: change to simple calculation

This project publishes no release notes. Between v0.4.7 and v0.4.8 there were 2 commits, 1 of them substantive:

• hostbuf-file-reader: TLS the ring buffer

This project publishes no release notes. Between v0.4.6 and v0.4.7 there were 5 commits, 4 of them substantive:

• model_vbar: make vbars_free_for_vbar lazy sync
• model_vbar: budget_deficit signed
• model_vbar: add surplus/deficit facility to vbars_free_for_vbar
• model_vbar: implement miss allocation check