Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

adding default inductor config settings #423

Merged
merged 13 commits into from
Jun 25, 2024
Merged

adding default inductor config settings #423

merged 13 commits into from
Jun 25, 2024

Conversation

HDCharles
Copy link
Contributor

@HDCharles HDCharles commented Jun 22, 2024 •
edited
Loading

Summary:

making autoquant and quantize (and eval and generate) apis call a new
recommended_inductor_config_setter util to set recommended apis

also update groupsize -> group_size in generate.py
and handled errors in autoquant (to pass CI)

Test Plan:

sh benchmarks.sh

comparison of different config combinations for matmul precision,
mixed_mm and coordinate_descent

high precision

tok/s= 9.14, mem/s= 60.55 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=147.02, mem/s= 973.53 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

medium precision

tok/s= 9.23, mem/s= 61.11 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=139.59, mem/s= 924.33 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

high + mixed_mm_choice heuristic

tok/s= 9.10, mem/s= 60.26 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=146.98, mem/s= 973.23 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

high + false use_mixed_mm

tok/s= 9.28, mem/s= 61.48 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=146.90, mem/s= 972.73 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

high + default mixed_mm_choice

tok/s= 9.08, mem/s= 60.09 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=137.58, mem/s= 911.00 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

high + heuristic + coordinate_descent_check_all_directions

tok/s= 9.19, mem/s= 60.87 GB/s, peak_mem= 8.61 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=166.02, mem/s=1099.30 GB/s, peak_mem= 8.97 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

high + false use_mixed_mm + coordinate_descent_check_all_directions

tok/s= 9.28, mem/s= 61.46 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=161.66, mem/s=1070.43 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

Reviewers:

Subscribers:

Tasks:

Tags:

@pytorch-bot PyTorch Bot
Copy link

pytorch-bot bot commented Jun 22, 2024 •
edited
Loading

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/423

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit d105072 with merge base 96d49cd (image):
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@HDCharles HDCharles requested a review from msaroufim June 22, 2024 18:30
@facebook-github-bot 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 Jun 22, 2024
msaroufim
msaroufim requested changes Jun 22, 2024
View reviewed changes
Copy link
Member

@msaroufim msaroufim left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice! Mind also updating any relevant documentation pages? Also if some of those flags are GPU specific just gate those behind a cuda flag

torchao/quantization/utils.py Show resolved Hide resolved
torchao/quantization/utils.py Show resolved Hide resolved
msaroufim
msaroufim approved these changes Jun 24, 2024
View reviewed changes
Copy link
Member

@msaroufim msaroufim left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cool LGTM

@HDCharles HDCharles force-pushed the 066_set_inductor_config branch 2 times, most recently from 5ab0b95 to bfe2ea2 Compare June 25, 2024 03:13
@HDCharles
Copy link
Contributor Author

Nice! Mind also updating any relevant documentation pages? Also if some of those flags are GPU specific just gate those behind a cuda flag

i don't think anything is gpu specific

@HDCharles
adding default inductor config settings
0e5fc3e 
Summary:

making autoquant and quantize apis call a new
recommended_inductor_config_setter util to set recommended apis

also update groupsize -> groupsize in generate.py

Test Plan:

sh benchmarks.sh

comparison of different config combinations for matmul precision,
mixed_mm and coordinate_descent

tok/s=  9.14, mem/s=  60.55 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=147.02, mem/s= 973.53 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.23, mem/s=  61.11 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=139.59, mem/s= 924.33 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.10, mem/s=  60.26 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=146.98, mem/s= 973.23 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.28, mem/s=  61.48 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=146.90, mem/s= 972.73 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.08, mem/s=  60.09 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=137.58, mem/s= 911.00 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.19, mem/s=  60.87 GB/s, peak_mem= 8.61 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=166.02, mem/s=1099.30 GB/s, peak_mem= 8.97 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
fixing tests
1421de8 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
fix weight only failures
c9932ce 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
fixing new broken test
18cd1aa 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
fixing autoquant test
ee5183f 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
testing if inductor config is the issue
d0e822b 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
are inductor configs somehow being set?
a5654f3 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
when is coordinate descent tuning beinng enabled?
26d9110 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
reset inductor config for tests
762ef41 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
more test fixes
b0c4e23 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
adding warning
8fadc39 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
handling of errors
3c2825e 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
@HDCharles
option to supress autoquant errors
d105072 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
jerryzh168
jerryzh168 reviewed Jun 25, 2024
View reviewed changes
test/integration/test_integration.py
@@ -689,6 +696,7 @@ def test_int8_dynamic_quant_subclass(self, device, dtype):

@parameterized.expand(COMMON_DEVICE_DTYPE)
def test_int8_weight_only_quant_subclass(self, device, dtype):
undo_recommended_configs()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why do we need these?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

a bunch of resource usage errors

https://github.com/pytorch/ao/actions/runs/9657170605/job/26635939215

@HDCharles HDCharles merged commit 211b6bc into main Jun 25, 2024
13 checks passed
dbyoung18 pushed a commit to dbyoung18/ao that referenced this pull request Jul 31, 2024
@HDCharles
adding default inductor config settings (pytorch#423)
7b03ef3 
* adding default inductor config settings

Summary:

making autoquant and quantize apis call a new
recommended_inductor_config_setter util to set recommended apis

also update groupsize -> groupsize in generate.py

Test Plan:

sh benchmarks.sh

comparison of different config combinations for matmul precision,
mixed_mm and coordinate_descent

tok/s=  9.14, mem/s=  60.55 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=147.02, mem/s= 973.53 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.23, mem/s=  61.11 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=139.59, mem/s= 924.33 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.10, mem/s=  60.26 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=146.98, mem/s= 973.23 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.28, mem/s=  61.48 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=146.90, mem/s= 972.73 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.08, mem/s=  60.09 GB/s, peak_mem= 8.33 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=137.58, mem/s= 911.00 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,
tok/s=  9.19, mem/s=  60.87 GB/s, peak_mem= 8.61 GB, model_size= 6.62 GB quant: int8dq, mod: Llama-2-7b-chat-hf,
tok/s=166.02, mem/s=1099.30 GB/s, peak_mem= 8.97 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf,

Reviewers:

Subscribers:

Tasks:

Tags:

* fixing tests

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* fix weight only failures

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* fixing new broken test

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* fixing autoquant test

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* testing if inductor config is the issue

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* are inductor configs somehow being set?

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* when is coordinate descent tuning beinng enabled?

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* reset inductor config for tests

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* more test fixes

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* adding warning

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* handling of errors

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* option to supress autoquant errors

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@jerryzh168 jerryzh168 jerryzh168 left review comments

@msaroufim msaroufim msaroufim approved these changes

Assignees
No one assigned
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@HDCharles @msaroufim @jerryzh168 @facebook-github-bot