Fused HQQ Quantization Gemm #153

jeromeku · 2024-04-22T02:53:05Z

Fused `int4 / fp16` Quant Matmul

Fused kernel that combines asymmetric dequantization and gemm. Useful primarily for compute-bound (M > 16) scenarios and not for memory-bound / inference scenarios.

The kernel fuses two ops:

Dequantization: upcasts u4 / s4 weights to float16 / bfloat16, followed by groupwise scaling and shifting by scales / zeropoints
GEMM: matmul on dequantized weights and activations.

Tested and benchmarked for HQQ but could theoretically be used for any asymmetric quantization scheme.

NOTE: Benchmark below is only indicative of performance on consumer-grade Ampere GPUs (A6000 specifically). When tested on H100, the performance is on par / marginally worse than native / compiled torch.
The intended use is thus for fine-tuning / training models on non-datacenter GPUs (80 <= compute capability < 90). If interested in optimizing the kernel for other architectures, please drop a note in the CUDA-MODE Discord channel.

Implementation Details

Bitpacking is simple row interleave, no need for extensive preprocessing (e.g., tinygemm or fastertransformer)
Tested for float16 / bfloat16 activations, scales, and zeros
Autotuned for both compute-bound and memory-bound configs
Assumes operand B of the gemm is is the quantized type.
Requires quantization along in-features, i.e., the K dimension, or axis=1, of torch.linear.weight.
Implementation handles both transposed and non-tranposed quantized weights, useful for forward / backward training passes.

Performance

Initial benchmarking (on A6000) demonstrates promising results, scaling well for compute-bound workloads:

	M	N	K	group_size	dtype	hqq_ref	triton	tinygemm
0	16	4096	4096	128	torch.bfloat16	0.2675	0.0633	0.0382
1	32	4096	4096	128	torch.bfloat16	0.2669	0.0704	0.0649
2	128	4096	4096	128	torch.bfloat16	0.2689	0.0960	0.2523
3	256	4096	4096	128	torch.bfloat16	0.3268	0.1355	0.5192
4	512	4096	4096	128	torch.bfloat16	0.3628	0.2369	1.0892
5	1024	4096	4096	128	torch.bfloat16	0.5133	0.4753	2.2016

Times are in ms, see benchmarks/benchmark_hqq.py.
hqq_ref is the base HQQ_Linear module that is unfused (dequantization followed by call to torch.matmul).
tinygemm calls torch.ops.aten._weight_int4pack_mm. Implementation is a custom HQQLinear layer that wraps the preprocessing necessary for this kernel, adapted from a benchmark script posted by @mobicham from CUDA-mode Discord discussions.

GPU details:

_CudaDeviceProperties(name='NVIDIA RTX A6000', major=8, minor=6, total_memory=48676MB, multi_processor_count=84)

NOTE

This implementation requires triton >= 3.0.0.

Running tests / benchmarks requires installation of hqq:
```
pip install hqq
```

TODO

clean up / refactor test to use pytest
add transposed matmul
torch.compile benchmarking
introduce additional prologue / epilogue fusions
create standalone dequant kernel
test across range of gpus
adapt for other quant methods (other than hqq)

msaroufim

first pass

msaroufim · 2024-04-22T16:20:40Z

benchmarks/benchmark_hqq.py

@@ -0,0 +1,134 @@
+import torch
+from termcolor import colored


will need to get rid of this dependency for merge, I'm fine with adding adding colors though so something like this should work

RED = "\033[31m" GREEN = "\033[32m" YELLOW = "\033[33m" BLUE = "\033[34m" MAGENTA = "\033[35m" CYAN = "\033[36m" WHITE = "\033[37m" RESET = "\033[0m" # Resets the color to default. name = "Alice" print(f"{GREEN}Hello, {name}!{RESET}")

msaroufim · 2024-04-22T16:21:11Z

torchao/prototype/hqq/mixed_mm.py

+from triton import cdiv
+import triton.language as tl
+from .kernels import mixed_mm_kernel_compute_bound, mixed_mm_kernel_max_autotune
+#credit jlebar


could you link to the original code as well? Will need to double check if the LICENSE allows us to copy paste

msaroufim · 2024-04-22T16:24:54Z

test/hqq/test_triton_mm.py

@@ -0,0 +1,101 @@
+import itertools


both test and benchmark will require skips if triton is less than 3.0 (which is fine because nightlies now ship with 3.0.0) and if hqq is not installed

For hqq I'm fine if we add it as a dev dependency for now

msaroufim · 2024-04-22T16:26:06Z

benchmarks/benchmark_hqq.py

+    return ref_time, tt_time, int4_time if dtype == torch.bfloat16 else None
+
+
+SHAPES = [


@cpuhrsch I guess these shapes are fine for now but are there some specific shapes we're more interested in tracking on an ongoing basis if so I wish we could just make them part of our benchmark or test utilities

msaroufim · 2024-04-22T16:28:30Z

benchmarks/benchmark_hqq.py

+
+
+df = pd.DataFrame(data, columns=HEADERS)
+df.to_csv("benchmark_triton.csv", index=False)


we will lose this csv on CI unless its saved to some github artifact so unless this file is huge let's just print it for now

msaroufim · 2024-04-22T16:48:37Z

torchao/prototype/hqq/README.md

+GPU details:
+
+```
+_CudaDeviceProperties(name='NVIDIA RTX A6000', major=8, minor=6, total_memory=48676MB, multi_processor_count=84)


once we figure out the installation issues I'll check to see if results repro on an H100

apologies meant pip freeze

pip_freeze.txt

@msaroufim

When you run on H100, can you run once with DISABLE_MMA_V3=1? It toggles Hopper specific specializations in triton. Curious to see how performance changes.

msaroufim · 2024-04-22T16:51:59Z

torchao/prototype/hqq/README.md

@@ -0,0 +1,43 @@
+## Fused `int4 / fp16` Quant Matmul
+
+Fused gemm for asymmetric quantized weights. Tested and benchmarked for `HQQ` but could theoretically be used for any asymmetric quantization scheme.


sounds like one of the 2 asymetric should be a symetric?

msaroufim · 2024-04-22T16:54:32Z

torchao/prototype/hqq/README.md

+The kernel packs `u4 / s4` weights and fuses dequantization with the matmul.
+
+- tested for `float16 / bfloat16` activations, scales, and zeros
+- autotuned for both compute-bound and io-bound configs


nit: could use memory bandwidth bound terminology instead

msaroufim · 2024-04-22T17:01:42Z

torchao/prototype/hqq/README.md

+
+Fused gemm for asymmetric quantized weights. Tested and benchmarked for `HQQ` but could theoretically be used for any asymmetric quantization scheme.
+
+The kernel packs `u4 / s4` weights and fuses dequantization with the matmul.


n00b q: whu can't we generically do this with torch.compile @HDCharles

It does work with torch.compile and there's a good speed-up (up to 4x compared to Pytorch), but a dequantize() CUDA kernel + torch.matmul is a bit faster.
I think the bitpacking should be done in such a way that torch.compile can fully optimize it.

msaroufim · 2024-04-22T17:04:30Z

torchao/prototype/hqq/kernels.py

+
+
+@triton.jit
+def _mixed_mm_kernel(


I see a lot of similarities between this code and what you had contributed for galore, can we start modularizing?

msaroufim

first pass

jeromeku · 2024-04-22T23:02:27Z

@msaroufim

Added a transpose implementation for backwards passes
Cleaned up test
- skips unless triton >= 3.0 + hqq
- refactored to use pytest
- added transpose test
- removed colored outputs
Benchmark
- cleaned up imports
- removed colors
- print benchmark results to terminal rather than saving as csv
Attached my pip env (see reply above). TLDR: torch 2.2.2, triton 3.0.0

msaroufim · 2024-04-24T23:51:27Z

Results on an H100

tinygemm degrades substantially relative to your triton implementation as the sizes increase
Results are always worst than the hqq implementation

Although it's very curious how different the pattern is on an A6000

(x) [[email protected] ~/ao (hqq_mixed_gemm)]$ python benchmarks/benchmark_hqq.py 
_CudaDeviceProperties(name='NVIDIA H100', major=9, minor=0, total_memory=97320MB, multi_processor_count=132)
shape=[16, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1194 Triton: 0.1678 Torch int4mm: 0.0248

shape=[32, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1203 Triton: 0.1567 Torch int4mm: 0.0442

shape=[128, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1253 Triton: 0.1607 Torch int4mm: 0.1564

shape=[256, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1269 Triton: 0.1515 Torch int4mm: 0.3025

shape=[512, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1344 Triton: 0.1477 Torch int4mm: 0.5926

shape=[1024, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1499 Triton: 0.1585 Torch int4mm: 1.1743

M,N,K,group_size,dtype,ref,triton,tinygemm
16,4096,4096,128,torch.bfloat16,0.11943938583135605,0.1677887886762619,0.024766016751527786
32,4096,4096,128,torch.bfloat16,0.12034595012664795,0.15668845176696777,0.044162582606077194
128,4096,4096,128,torch.bfloat16,0.1253334879875183,0.16072532534599304,0.15640245378017426
256,4096,4096,128,torch.bfloat16,0.1268536001443863,0.1515069156885147,0.3024669885635376
512,4096,4096,128,torch.bfloat16,0.13444344699382782,0.14771266281604767,0.5926073789596558
1024,4096,4096,128,torch.bfloat16,0.14986468851566315,0.15848150849342346,1.1743202209472656

(x) [[email protected] ~/ao (hqq_mixed_gemm)]$ DISABLE_MMA_V3=1 python benchmarks/benchmark_hqq.py 
_CudaDeviceProperties(name='NVIDIA H100', major=9, minor=0, total_memory=97320MB, multi_processor_count=132)
shape=[16, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1192 Triton: 0.1538 Torch int4mm: 0.0248

shape=[32, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1202 Triton: 0.1587 Torch int4mm: 0.0441

shape=[128, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1253 Triton: 0.1532 Torch int4mm: 0.1567

shape=[256, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1264 Triton: 0.1610 Torch int4mm: 0.3028

shape=[512, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1334 Triton: 0.1618 Torch int4mm: 0.5928

shape=[1024, 4096, 4096] group_size=128 dtype=torch.bfloat16:
Ref: 0.1496 Triton: 0.1632 Torch int4mm: 1.1754

M,N,K,group_size,dtype,ref,triton,tinygemm
16,4096,4096,128,torch.bfloat16,0.11921614408493042,0.15375947952270508,0.02484913170337677
32,4096,4096,128,torch.bfloat16,0.12021129578351974,0.15867114067077637,0.04411611706018448
128,4096,4096,128,torch.bfloat16,0.12525483965873718,0.15317335724830627,0.1566852480173111
256,4096,4096,128,torch.bfloat16,0.12637671828269958,0.16101239621639252,0.30275848507881165
512,4096,4096,128,torch.bfloat16,0.1334434598684311,0.16184988617897034,0.5928144454956055
1024,4096,4096,128,torch.bfloat16,0.14962367713451385,0.16317297518253326,1.1754025220870972

msaroufim

Thank you!

* add test / benchmark * add kernels * update readme * more readme edits * edit readme * add transpose test * transpose test pass * refactor test * add checks for CI * add more comments for transpose kernel * remove import in test * clean up benchmark * fix test import order * minor README edits * additional readme edits * update readme * update readme * add note about cudamode --------- Co-authored-by: Mark Saroufim <[email protected]>

jeromeku added 4 commits April 22, 2024 02:34

add test / benchmark

0d41668

add kernels

497d8db

update readme

8db9f51

more readme edits

2a18357

facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Apr 22, 2024

jeromeku marked this pull request as draft April 22, 2024 02:53

edit readme

f11a59f

msaroufim reviewed Apr 22, 2024

View reviewed changes

msaroufim requested changes Apr 22, 2024

View reviewed changes

jeromeku added 7 commits April 22, 2024 19:06

add transpose test

2e76839

transpose test pass

19c43c2

refactor test

e0f3781

add checks for CI

be718d6

add more comments for transpose kernel

793994b

remove import in test

5c585b9

clean up benchmark

c80b239

jeromeku added 4 commits April 22, 2024 23:17

fix test import order

48a153f

minor README edits

b3a9ab8

additional readme edits

bbe9083

update readme

38dbc3e

jeromeku marked this pull request as ready for review April 24, 2024 20:19

msaroufim self-requested a review April 24, 2024 23:15

jeromeku added 2 commits April 25, 2024 02:00

update readme

d89fa74

add note about cudamode

cd68d38

msaroufim approved these changes Apr 25, 2024

View reviewed changes

Merge branch 'main' into hqq_mixed_gemm

582fd8b

msaroufim merged commit e148244 into pytorch:main Apr 25, 2024
13 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fused HQQ Quantization Gemm #153

Fused HQQ Quantization Gemm #153

jeromeku commented Apr 22, 2024 •

edited

Loading

msaroufim left a comment

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024 •

edited

Loading

jeromeku Apr 22, 2024

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024

jeromeku Apr 22, 2024

jeromeku Apr 24, 2024 •

edited

Loading

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024

msaroufim Apr 22, 2024

mobicham Apr 22, 2024

msaroufim Apr 22, 2024

msaroufim left a comment

jeromeku commented Apr 22, 2024

msaroufim commented Apr 24, 2024 •

edited

Loading

msaroufim left a comment

		return ref_time, tt_time, int4_time if dtype == torch.bfloat16 else None


		SHAPES = [



		df = pd.DataFrame(data, columns=HEADERS)
		df.to_csv("benchmark_triton.csv", index=False)

		@@ -0,0 +1,43 @@
		## Fused `int4 / fp16` Quant Matmul

		Fused gemm for asymmetric quantized weights. Tested and benchmarked for `HQQ` but could theoretically be used for any asymmetric quantization scheme.


		Fused gemm for asymmetric quantized weights. Tested and benchmarked for `HQQ` but could theoretically be used for any asymmetric quantization scheme.

		The kernel packs `u4 / s4` weights and fuses dequantization with the matmul.



		@triton.jit
		def _mixed_mm_kernel(

Fused HQQ Quantization Gemm #153

Fused HQQ Quantization Gemm #153

Conversation

jeromeku commented Apr 22, 2024 • edited Loading

Fused int4 / fp16 Quant Matmul

Implementation Details

Performance

NOTE

TODO

msaroufim left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

msaroufim Apr 22, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jeromeku Apr 24, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

msaroufim left a comment

Choose a reason for hiding this comment

jeromeku commented Apr 22, 2024

msaroufim commented Apr 24, 2024 • edited Loading

msaroufim left a comment

Choose a reason for hiding this comment

jeromeku commented Apr 22, 2024 •

edited

Loading

Fused `int4 / fp16` Quant Matmul

msaroufim Apr 22, 2024 •

edited

Loading

jeromeku Apr 24, 2024 •

edited

Loading

msaroufim commented Apr 24, 2024 •

edited

Loading