[inductor] initial triton static config lookup table #157699

coconutruben · 2025-07-07T08:46:00Z

Stack from ghstack (oldest at bottom):

Summary:

Why

enable initial feature set to see wider internal benchmarking and adoption
introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on

What

First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm
supports triton, tma, decompose_k, and bias_addmm in configs
configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton

Test Plan:

buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu
buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu
buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e

Rollback Plan:

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov

Differential Revision: D77895791

Summary: \# Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on \# What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: [ghstack-poisoned]

pytorch-bot · 2025-07-07T08:46:04Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/157699

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

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

✅ No Failures

As of commit 0895c76 with merge base ecea811 ():
💚 Looks good so far! There are no failures yet. 💚

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

coconutruben · 2025-07-07T08:49:17Z

apologies @jansel @masnesral @PaulZhang12 I broke this out from #156785 to be github first and speed up linting and development etc. This version addresses some of the feedback from last week, namely having a way of passing through EVEN_K and ALLOW_TF32

please lmk what else we should address here

Summary: \# Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on \# What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]

coconutruben · 2025-07-07T22:08:02Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

torch/_inductor/lookup_table.py

Summary: \# Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on \# What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-08T00:17:42Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: \# Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on \# What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-08T02:31:36Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: \# Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on \# What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-08T18:24:13Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

coconutruben · 2025-07-09T00:26:18Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

torch/_inductor/config.py

torch/_inductor/kernel/bmm.py

jansel · 2025-07-09T05:05:06Z

torch/_inductor/kernel/mm_common.py

+        assert len(input_nodes) == 3 and input_nodes[0] == mat
+        size = V.graph.sizevars.size_hints(
+            size,
+            fallback=torch._inductor.config.unbacked_symint_fallback,


Are we hitting unbacked symints in practice? I worry the tuned configs in these causes could be false positives since we don't have realistic shapes. Maybe we should not use the lookup table in this case?

I haven't seen it in practice yet, but I also haven't looked too thoroughly. If there is an unbacked symint in practice, what happens in the current autotuning logic i.e. how does the sample input get generated for benchmarking it? we can skip the table, but we can also just match whatever that does, as we won't be worse?

I'm not sure, let's double check.

https://github.com/pytorch/pytorch/blob/main/torch/_inductor/autotune_process.py#L341
TensorMeta uses the fallback as well to generate the sizes and strides, and we use those to generate tensors for benchmarking - so it seems it's the same as we're doing here

torch/_inductor/lookup_table.py

torch/_inductor/template_heuristics.py

etaf · 2025-07-09T06:17:05Z

Added ciflow/xpu to check for potential XPU breakages. Apologies for any inconvenience caused.

torch/_inductor/config.py

torch/_inductor/template_heuristics.py

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-23T19:27:14Z

ALLOW_TF32 should be part of the key since it changes the semantics of the program.

What difference do you see between having ALLOW_TF32 as part of the key and having it as part of the retrieved value, but we filter out all the values that don't match?
The reason it's in the value is because we need it for template instantiation. If we add it to the key, are we fine keeping it inside the key and the value?

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-23T19:35:09Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

coconutruben · 2025-07-23T19:40:40Z

ALLOW_TF32 should be part of the key since it changes the semantics of the program.

asked differently, @jansel do you just want to have the value of torch.backends.cuda.matmul.allow_tf32 inside the key?

jansel · 2025-07-23T20:00:04Z

asked differently, @jansel do you just want to have the value of torch.backends.cuda.matmul.allow_tf32 inside the key?

Yes exactly.

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-23T23:07:02Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

coconutruben · 2025-07-23T23:07:52Z

torch.backends.cuda.matmul.allow_tf32

it's now part of the key, and we have this general notion of "input_key_suffix" where we can stuff inductor wide things into that we want to hold

this is a new style input key

"NVIDIA H100+mm+((torch.bfloat16, [1024, 1024], [1024, 1]), (torch.bfloat16, [1024, 1024], [1024, 1]))+tf32=False"

coconutruben · 2025-07-24T06:38:39Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-25T08:43:36Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-25T16:40:30Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-28T09:26:08Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-07-29T17:40:18Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Summary: # Why - enable initial feature set to see wider internal benchmarking and adoption - introduce expected behavior testing that subsequent expansions (more backends, functions, etc) can rely on # What - First version of a static lookup table for Triton configs across mm, addmm, bmm, mm_plus_mm - supports triton, tma, decompose_k, and bias_addmm in configs - configuration inside lookup_table.py for now, with knob to turn on/off in inductor_config.triton Test Plan: ``` buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:template_heuristics_cpu buck2 test mode/opt fbcode//caffe2/test/inductor:lookup_table_e2e ``` Rollback Plan: cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov Differential Revision: [D77895791](https://our.internmc.facebook.com/intern/diff/D77895791) [ghstack-poisoned]

coconutruben · 2025-08-11T11:01:28Z

@coconutruben has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

coconutruben · 2025-08-11T11:02:40Z

torch/_inductor/kernel/mm.py

@@ -319,13 +320,13 @@

        {%- if TMA_EXPERIMENTAL_API %}
        a = tl._experimental_descriptor_load(
-            a_desc,


@PaulZhang12 can you double check if this is right and makes sense? after rebasing now I kept failing on tma with TMA_EXPERIMENTAL_API, and it seems to me the code here is just wrong (a_desc is not defined when running without TMA_EXPERIMENTAL_API)

eellison

Instead of special casing the config lookup table, can we make the lookup table part of a generic api to override the config selection ? '

We should be able to use one code path for both the config lookup table, and the prediction model.

coconutruben · 2025-08-11T21:59:07Z

Instead of special casing the config lookup table, can we make the lookup table part of a generic api to override the config selection ? '

We should be able to use one code path for both the config lookup table, and the prediction model.

@eellison I can make the lookup table a more generic "config overrider" interface/class, but I'm wondering, what do you and @jansel / @exclamaforte think about treating the performance model as a (generic) config filter in this case? i.e. the interplay of standard config generation, lookup table (LUT) and performance model being

LUT or config heuristic generate the configs
performance models filters them down to topk

I think this would be better, because the performance model is at its most useful when it just rates (runs a prediction) on each config, and then picks the topk from that, rather than the contract being "generate me the best configs"

I plan to evolve the get_mm_configs in choices to a more generic get_template_configs, and take in the full list of templates (in use e.g. triton mm, triton tma, cutlass mm) for one op, iterate over those, and generate all the configs/choices. This then allows us to do one op wide override from the lookup table (just get all the template/config pairs for that op/input from the table) and have the performance model run eval on all the configs (or the ones it can). This also allows op-wide decisions in a single place (e.g. are there 0 matches in the table, do we fallback to ATEN), etc.

what do you guys think of that approach, of having a config generator interface (config heuristics, lookup table) and a config filter interface (performance model)?

jansel · 2025-08-12T03:19:42Z

I want both of these to go throw a common extension point in choices.py.

pytorch-bot bot added ciflow/inductor module: inductor labels Jul 7, 2025

coconutruben mentioned this pull request Jul 7, 2025

wip, lookup table for reduction configs #157700

Open

coconutruben requested review from jansel, PaulZhang12 and masnesral July 7, 2025 08:46

coconutruben added the topic: not user facing topic category label Jul 7, 2025

pytorch-bot bot added the ciflow/trunk Trigger trunk jobs on your pull request label Jul 7, 2025

coconutruben commented Jul 7, 2025

View reviewed changes

torch/_inductor/lookup_table.py Outdated Show resolved Hide resolved

coconutruben mentioned this pull request Jul 8, 2025

[inductor][lookup_table] log entries #157768

Open

jansel requested changes Jul 9, 2025

View reviewed changes

etaf added the ciflow/xpu Run XPU CI tasks label Jul 9, 2025

PaulZhang12 reviewed Jul 9, 2025

View reviewed changes

torch/_inductor/config.py Outdated Show resolved Hide resolved

PaulZhang12 reviewed Jul 9, 2025

View reviewed changes

torch/_inductor/template_heuristics.py Outdated Show resolved Hide resolved

coconutruben mentioned this pull request Jul 23, 2025

[inductor] add lookup table recorder #158987

Open

coconutruben added 2 commits July 29, 2025 08:39

coconutruben mentioned this pull request Jul 29, 2025

[inductor] consolidate common GEMM triton param retrieval #159383

Closed

coconutruben requested a review from eellison July 30, 2025 04:14

BoyuanFeng self-requested a review August 9, 2025 03:36

coconutruben commented Aug 11, 2025

View reviewed changes

eellison reviewed Aug 11, 2025

View reviewed changes

[inductor] initial triton static config lookup table #157699

Are you sure you want to change the base?

[inductor] initial triton static config lookup table #157699

Uh oh!

Conversation

coconutruben commented Jul 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Why

What

Uh oh!

pytorch-bot bot commented Jul 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/157699

✅ No Failures

Uh oh!

coconutruben commented Jul 7, 2025

Uh oh!

coconutruben commented Jul 7, 2025

Uh oh!

Uh oh!

coconutruben commented Jul 8, 2025

Uh oh!

coconutruben commented Jul 8, 2025

Uh oh!

coconutruben commented Jul 8, 2025

Uh oh!

coconutruben commented Jul 9, 2025

Uh oh!

Uh oh!

Uh oh!

jansel Jul 9, 2025

Choose a reason for hiding this comment

Uh oh!

coconutruben Jul 9, 2025

Choose a reason for hiding this comment

Uh oh!

jansel Jul 9, 2025

Choose a reason for hiding this comment

Uh oh!

coconutruben Jul 9, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

etaf commented Jul 9, 2025

Uh oh!

Uh oh!

Uh oh!

coconutruben commented Jul 23, 2025

Uh oh!

coconutruben commented Jul 23, 2025

Uh oh!

coconutruben commented Jul 23, 2025

Uh oh!

jansel commented Jul 23, 2025

Uh oh!

coconutruben commented Jul 23, 2025

Uh oh!

coconutruben commented Jul 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coconutruben commented Jul 24, 2025

Uh oh!

coconutruben commented Jul 25, 2025

Uh oh!

coconutruben commented Jul 25, 2025

Uh oh!

coconutruben commented Jul 28, 2025

Uh oh!

coconutruben commented Jul 29, 2025

Uh oh!

coconutruben commented Aug 11, 2025

Uh oh!

coconutruben Aug 11, 2025

Choose a reason for hiding this comment

Uh oh!

coconutruben commented Jul 7, 2025 •

edited

Loading

pytorch-bot bot commented Jul 7, 2025 •

edited

Loading

coconutruben commented Jul 23, 2025 •

edited

Loading