Feat/mps gemm 4bit bf16 - #2070

Closed

eaglstun wants to merge 23 commits into

bitsandbytes-foundation:mainfrom

eaglstun:feat/mps-gemm-4bit-bf16

Commits on Jul 8, 2026

Phase 1: MPS correctness sweep — parity harness + ground-truth audit

eaglstun
and
claude
committed
0273608
Fix Lion to use decoupled weight decay in the default backend

eaglstun
and
claude
committed
5f37b8b
Phase 2: quantize_blockwise through hand-written Metal (native pipe, one op end to end)

eaglstun
and
claude
committed
cbfa1ad
Phase 3: dequantize_blockwise + dequantize_4bit + quantize_4bit on native Metal, and a load-time buffer-contract guard

eaglstun
and
claude
committed
863990c
Merge MPS native Metal backend (Phases 1-3): quantize/dequantize blockwise + 4-bit

eaglstun
and
claude
committed
501585f
Packaging: ship bitsandbytes.metallib in the wheel so native MPS works from pip install

eaglstun
and
claude
committed
705a6e8
Phase 4: Apple Silicon / MPS user-facing docs + next-phase matmul plan

eaglstun
and
claude
committed
642ecb1

Commits on Jul 15, 2026

Phase M1: record 4-bit matmul baseline + A/B design decision

eaglstun
and
claude
committed
4598cd2
Phase M2: fused native gemv_4bit Metal kernel (dequant + dot product, no B_dq)

eaglstun
and
claude
committed
d913182
Phase M3: native gemm_4bit -- chunked dequant + MPSMatrixMultiplication, one command buffer

eaglstun
and
claude
committed
bdd3d9e
Phase M4: matmul closeout -- sync tax measured + documented, offset-0 semantics verified, honest docs

eaglstun
and
claude
committed
0c9db70
Merge MPS 4-bit matmul sub-phase (M1-M4): native gemv + gemm on Metal

eaglstun
and
claude
committed
ed2ea2e
Retire the lion weight-decay strict-xfail now that the divergence is fixed

eaglstun
and
claude
committed
5e7ff4e
Fix Lion to use decoupled weight decay in the Triton 32-bit kernel

eaglstun
and
claude
committed
45813c3
Fix Lion (and Adagrad) weight decay in the Triton 8-bit blockwise kernel

eaglstun
and
claude
committed
7333266
Fix Lion to use decoupled weight decay in the CUDA 32-bit kernel

eaglstun
and
claude
committed
6a81fa9
Merge pull request #1 from eaglstun/mps-lion-triton-32bit
eaglstun
authored
7aa17ce
Merge pull request #2 from eaglstun/mps-lion-triton-8bit
eaglstun
authored
0c86556
Merge pull request #3 from eaglstun/mps-lion-cuda
eaglstun
authored
9afd664
Merge upstream/main into fork main (Apple Silicon MPS line)

eaglstun
and
claude
committed
d29189f

Commits on Aug 28, 2026

Merge remote-tracking branch 'upstream/main'
eaglstun
committed
680ad87
Migrate dispatch tooling to agent configuration
eaglstun
committed
7582cb3

Commits on Sep 4, 2026

mps 4bit bf16
eaglstun
committed
0ddb386