IFU master 2024 01 24 - Githubissues

ROCm / bitsandbytes

8-bit CUDA functions for PyTorch

MIT License

34 stars 3 forks source link

IFU master 2024 01 24 #4

Closed pnunna93 closed 8 months ago

pnunna93 commented 8 months ago

This PR pulls upstream changes for 0.42.0 version.

Resolved merge conflicts - conflicts_diff.txt
Updated hipified files for new kernels and ops from upstream
Fixed build errors and ran unittests, summary shown below
Skipped failing unit tests for future review. There are around 200 additional tests added in functional module, which contributed to major chunk of new failures.

Unit test summary: PreIFU:	Module	Passed	Failed	Skipped
autograd	1616	624	0
cuda_setup_evaluator	0	1	0
functional	270	23	43
linear8bitlt	9	9	0
modules	10	4	0
optim	125	26	26
triton	0	2	0
Total	2030	689	69

PostIFU:	Module	Passed	Failed	Skipped
autograd	1592	648	0
cuda_setup_evaluator	0	1	0
functional	313	233	54
linear8bitlt	9	9	0
modules	14	4	0
optim	124	27	26
triton	0	2	0
generation	8	8	0
linear4bit	32	0	0
Total	2092	932	80