Repository navigation

[OMNIML-5899] Add CUDA packing for IQ formats - #2445

Closed

hychiang-git wants to merge 33 commits into

hungyuehc/omniml-5899-ggmlfrom

hungyuehc/omniml-5899-cuda

Commits on Sep 16, 2026
- Add CUDA packing for IQ formats
  
  ChenhanYu
  authored and
  hychiang-git
  committed
  70bfa81
- Address IQ CUDA review feedback
  hychiang-git
  committed
  d830eb3
- Fix IQ CUDA staging and parity tests
  hychiang-git
  committed
  e8ab3cd
- Stabilize IQ2 CUDA scale checks
  hychiang-git
  committed
  ce64bd8
- Document IQ1 CUDA scale constants
  hychiang-git
  committed
  c227115
- Wire IQ CUDA packers into dispatch
  hychiang-git
  committed
  ac6b4b3
- Name IQ2 CUDA scale constants
  hychiang-git
  committed
  984be20
- Strengthen IQ CUDA dispatch coverage
  hychiang-git
  committed
  98a823a
- Assert IQ CUDA launch invariants
  hychiang-git
  committed
  72a0480
- Unify IQ1 CUDA error evaluation
  hychiang-git
  committed
  95b7a4c
- Derive IQ CUDA payload offsets
  hychiang-git
  committed
  e585650
- Pin IQ1 CUDA choice geometry
  hychiang-git
  committed
  2da89c9
- Strengthen IQ CUDA geometry checks
  hychiang-git
  committed
  7625e5a
- Harden IQ CUDA build and metadata checks
  hychiang-git
  committed
  4eb23b9
- Harden IQ CUDA bit-field validation
  hychiang-git
  committed
  8dc6c2e
- Strengthen IQ CUDA fallback coverage
  hychiang-git
  committed
  327be16
- Keep IQ CUDA extensions lazy
  hychiang-git
  committed
  ab1e808
- Restore explicit IQ extension precompile
  hychiang-git
  committed
  6035511
- Stabilize IQ fallback parity tests
  hychiang-git
  committed
  9458cc8
- Harden IQ CUDA validation and tests
  hychiang-git
  committed
  afeb635
- Expose reference IQ search and validate grids
  hychiang-git
  committed
  c28c16e
- Protect canonical IQ grid caches
  hychiang-git
  committed
  389f7c1
- Fold scale search into IQ CUDA encoders
  hychiang-git
  committed
  cfcf90a
- Keep fake tensors off IQ CUDA dispatch
  hychiang-git
  committed
  83bd02f
- Tighten IQ CUDA validation
  hychiang-git
  committed
  f6ad3c2
- Harden IQ CUDA review contracts
  hychiang-git
  committed
  c5038f2
- Honor IQ search changes across caches
  hychiang-git
  committed
  4cf8791
- Tighten IQ CUDA launch contracts
  hychiang-git
  committed
  83675c4
- Document IQ native packing contracts
  hychiang-git
  committed
  8726e7b
- Cache IQ extension build failures
  hychiang-git
  committed
  5bfda80
- Clarify block reduction contract
  hychiang-git
  committed
  84f7c09
- Align IQ CUDA input and packaging contracts
  hychiang-git
  committed
  067bdee
- Enforce native IQ CUDA dtype contracts
  hychiang-git
  committed
  0c81d1d

Navigation Menu

Repository navigation

[OMNIML-5899] Add CUDA packing for IQ formats - #2445

[OMNIML-5899] Add CUDA packing for IQ formats#2445
hychiang-git wants to merge 33 commits into
hungyuehc/omniml-5899-ggmlfrom
hungyuehc/omniml-5899-cuda

Commits on Sep 16, 2026

Add CUDA packing for IQ formats

Address IQ CUDA review feedback

Fix IQ CUDA staging and parity tests

Stabilize IQ2 CUDA scale checks

Document IQ1 CUDA scale constants

Wire IQ CUDA packers into dispatch

Name IQ2 CUDA scale constants

Strengthen IQ CUDA dispatch coverage

Assert IQ CUDA launch invariants

Unify IQ1 CUDA error evaluation

Derive IQ CUDA payload offsets

Pin IQ1 CUDA choice geometry

Strengthen IQ CUDA geometry checks

Harden IQ CUDA build and metadata checks

Harden IQ CUDA bit-field validation

Strengthen IQ CUDA fallback coverage

Keep IQ CUDA extensions lazy

Restore explicit IQ extension precompile

Stabilize IQ fallback parity tests

Harden IQ CUDA validation and tests

Expose reference IQ search and validate grids

Protect canonical IQ grid caches

Fold scale search into IQ CUDA encoders

Keep fake tensors off IQ CUDA dispatch

Tighten IQ CUDA validation

Harden IQ CUDA review contracts

Honor IQ search changes across caches

Tighten IQ CUDA launch contracts

Document IQ native packing contracts

Cache IQ extension build failures

Clarify block reduction contract

Align IQ CUDA input and packaging contracts

Enforce native IQ CUDA dtype contracts