[release/2.13] Advance Triton pin to 3.8.x branch tip

[naromero77amd]

Summary

• Advance the ROCm Triton pin to the tip of ROCm/triton release/internal/3.8.x.
• Update .ci/docker/triton_version.txt from 3.7.1 to 3.8.0.

Triton pin: 4cff872ced001ea92d9fcf05b3f6517e2b486d19
Triton branch: ROCm/triton@release/internal/3.8.x

Test plan

• git diff --check -- .ci/docker/ci_commit_pins/triton.txt .ci/docker/triton_version.txt
• Verified Triton checkout reports __version__ = '3.8.0'
• Run PyTorch CI

Follows the structure of #3451.

Made with Cursor

[naromero77amd]

Triton 3.8 bump fixes over 100 unit tests (mostly flex_attention) on MI450, more details here: https://github.com/ROCm/frameworks-internal/issues/17274#issuecomment-4986171237

NOTE: Results below are on MI300.

Adding the all-Inductor regression analysis from the Triton 3.8 validation run.

Triton 3.8 vs Triton 3.7 Regression Analysis

Baseline: Triton 3.7.1 all-Inductor result from the validation artifacts, using ROCm/triton release/internal/3.7.x commit 0263a6a6203cf27c441c57a6c808ea87ffb8f654.

Comparison: Triton 3.8.0 from ROCm/triton release/internal/3.8.x commit 4cff872ced001ea92d9fcf05b3f6517e2b486d19.

Important caveat: test/inductor/test_torchinductor_opinfo_properties.py was included in both compared result sets and completed normally in normal file mode. No host-side MemoryError, timeout, or missed tests were observed for this suite in either run.

Summary

Metric	Triton 3.7.1 baseline	Triton 3.8.0	Delta
Completed tests	27,732 / 27,732	27,732 / 27,732	0
Passed	22,404	22,402	-2
Skipped	4,912	4,912	+0
Xfailed	318	318	+0
Failed	95	97	+2
Failure rate	0.343%	0.350%	+0.007 pp
Error	3	3	+0
Timed out	0	0	+0
Missed	0	0	+0

xychart-beta
    title "Failed tests by Triton version"
    x-axis ["Triton 3.7.1", "Triton 3.8.0"]
    y-axis "Failed tests" 0 --> 120
    bar [95, 97]

Loading

pie showData
    title Triton 3.8 failure delta vs 3.7
    "Unchanged failures" : 98
    "New failures" : 2
    "Resolved failures" : 0

Loading

Interpretation

Triton 3.8.0 is very close to the Triton 3.7.1 baseline for the full all-Inductor suite on PyTorch 2.13. Total failures increased from 95 to 97, with errors unchanged at 3 and no timeouts or missed tests in either run. There are 2 new failing tests and 0 resolved failures.

Suite-Level Regressions

Suite	Triton 3.7.1 failed	Triton 3.8.0 failed	Delta
test/inductor/test_fused_attention.py	0	1	+1
test/inductor/test_padding.py	0	1	+1

New failing tests:

• test/inductor/test_fused_attention.py::SDPAPatternRewriterGpuTests::test_sdpa_rewriter_16_fp32_mask_gpu
• test/inductor/test_padding.py::PaddingTest::test_reduction_comprehensive_padding_stride

Suite-Level Improvements

No suite-level improvements.

Resolved tests:

• None

Unchanged Failing Suites

Suite	Triton 3.7.1 failed	Triton 3.8.0 failed	Delta
test/inductor/test_aot_inductor.py	2	2	+0
test/inductor/test_aot_inductor_arrayref.py	1	1	+0
test/inductor/test_aot_inductor_custom_ops.py	1	1	+0
test/inductor/test_aoti_torchbind_constants.py	3	3	+0
test/inductor/test_benchmarking.py	2	2	+0
test/inductor/test_ck_backend.py	18	18	+0
test/inductor/test_compile_subprocess.py	1	1	+0
test/inductor/test_compiled_autograd.py	27	27	+0
test/inductor/test_cpu_repro.py	2	2	+0
test/inductor/test_fp8.py	4	4	+0
test/inductor/test_gpu_cpp_wrapper.py	2	2	+0
test/inductor/test_inductor_utils.py	1	1	+0
test/inductor/test_layout_optim.py	5	5	+0
test/inductor/test_max_autotune.py	1	1	+0
test/inductor/test_mix_order_reduction.py	4	4	+0
test/inductor/test_pattern_matcher.py	4	4	+0
test/inductor/test_select_algorithm.py	1	1	+0
test/inductor/test_torchbind.py	16	16	+0
test/inductor/test_torchinductor.py	1	1	+0
test/inductor/test_torchinductor_dynamic_shapes.py	2	2	+0

Source validation thread: naromero77amd/framework_scripts#2 (comment)

Related draft PR: #3467

Local artifacts:

• triton37_log: /home/niromero/docker_workspace/pytorch_sandbox_3/inductor_all_pytorch_2_13_triton_3_7_20260721_220611.log
• triton37_analysis: /home/niromero/docker_workspace/pytorch_sandbox_3/inductor_all_pytorch_2_13_triton_3_7_20260721_220611.analysis.md
• triton38_log: /home/niromero/docker_workspace/pytorch_sandbox_3/inductor_all_pytorch_2_13_triton_3_8_20260722_152835.log
• triton38_analysis: /home/niromero/docker_workspace/pytorch_sandbox_3/inductor_all_pytorch_2_13_triton_3_8_20260722_152835.analysis.md
• comparison_md: /home/niromero/docker_workspace/pytorch_sandbox_3/triton_3_8_vs_3_7_pytorch_2_13_regression_20260723_143912.md
• comparison_json: /home/niromero/docker_workspace/pytorch_sandbox_3/triton_3_8_vs_3_7_pytorch_2_13_regression_20260723_143912.json

[jataylo]

lgtm

[naromero77amd]

Building with theROCk CI, the build succeeds
https://github.com/ROCm/TheRock/actions/runs/30285798625

this was only with the gfx94X-dcgpu.

Note that test_torchinductor.py is not included in the github actions workflow test matrix.