skip unsupported arches in python tests #5760

liqiangxl · 2026-01-05T16:25:16Z

skip unsupported arches in grouped and scaled matmuls

liqiangxl · 2026-01-05T16:25:29Z

!test

greptile-apps · 2026-01-05T16:28:11Z

Greptile Overview

Greptile Summary

Overview

This PR updates architecture checks in Python tests to skip unsupported GPU architectures for FP4 quantization and GEMM operations. The changes replace module-level skip conditions with per-test @pytest.mark.skipif decorators.

Key Changes

Added is_blackwell() helper in utils.py to identify Blackwell GPU variants
Replaced module-level skip in test_cutlass_nvfp4_gemm.py with per-test decorators
Consolidated skip conditions across test files to use microarchitecture_is(10, 0) for most tests
Simplified dual skip conditions (e.g., is_pre_blackwell() + microarchitecture_is_pre(12) → single check)

Critical Issues Found

1. Incorrect Architecture Definition in `is_blackwell()`

The new is_blackwell() function includes compute capabilities 12.0 (RTX PRO 6000/RTX 50XX) and 12.1 (DGX Spark) as Blackwell architectures. However, the cutlass scheduler (csrc/scheduler/cutlass.cpp:75-84) explicitly only supports compute capabilities 10.0 and 10.3:

if (device_prop->major != 10 || !(device_prop->minor == 0 || device_prop->minor == 3)) {
  // Error: "Cutlass scheduler only supports GB200 and GB300 (cc 10.0 or 10.3)"
}

Impact: Tests using is_blackwell() will incorrectly run on 12.0/12.1 devices and fail at runtime.

2. Inconsistent Skip Conditions in `test_narrow_precision.py`

test_scaled_mm uses microarchitecture_is(10, 0) - only 10.0
test_scaled_mm_nv_quantized uses is_blackwell() - includes unsupported 12.0/12.1
test_cutlass_nvfp4_grouped_mm uses microarchitecture_is(10, 0) - only 10.0

All these tests use nvfuser operations that may rely on cutlass scheduling. The inconsistency creates unclear test coverage and will cause failures on 12.0/12.1 devices for test_scaled_mm_nv_quantized.

3. More Restrictive Than Original Logic

The old skip conditions checked:

is_pre_blackwell() (major < 10) + not microarchitecture_is_pre(12) (major < 12)
Result: Ran on 10.0, 10.3, and any 11.x devices

The new condition microarchitecture_is(10, 0) only runs on exactly 10.0, excluding:

10.3 (B300/GB300) - which cutlass scheduler supports
11.x devices - unclear if these exist or were intended to be supported

Impact: Reduced test coverage, particularly for 10.3 which is explicitly supported by the cutlass scheduler.

Recommendations

Fix is_blackwell(): Remove 12.0 and 12.1, or rename the function to clarify it's broader than cutlass support
Standardize skip conditions: Decide whether to support only 10.0 or both 10.0 and 10.3, then apply consistently
Update test_scaled_mm_nv_quantized: Change to use microarchitecture_is(10, 0) or the corrected is_blackwell() to prevent runtime failures on unsupported architectures

Confidence Score: 2/5

This PR has critical logic errors that will cause test failures on compute capability 12.0/12.1 devices
Score of 2 reflects three critical issues: (1) is_blackwell() incorrectly includes unsupported compute capabilities 12.0 and 12.1, which will cause runtime failures when tests attempt to use cutlass operations on these devices; (2) inconsistent skip conditions across tests create unclear behavior and potential failures; (3) the new logic is more restrictive than the original, potentially excluding supported architectures like 10.3. While the module-level to per-test skip migration is correct, the implementation has significant logical flaws that need correction before merge.
tests/python/direct_utils/utils.py (fix is_blackwell definition) and tests/python/direct/test_narrow_precision.py (standardize skip conditions)

Important Files Changed

File Analysis

Filename	Score	Overview
tests/python/direct_utils/utils.py	3/5	Added is_blackwell() helper but incorrectly includes compute 12.0 and 12.1 which are not supported by cutlass scheduler
tests/python/direct/test_cutlass_nvfp4_gemm.py	5/5	Correctly replaced module-level skip with per-test skipif decorators using microarchitecture_is(10, 0)
tests/python/direct/test_narrow_precision.py	2/5	Inconsistent skip conditions: test_scaled_mm_nv_quantized uses is_blackwell() (includes unsupported 12.0/12.1) while other tests correctly use microarchitecture_is(10, 0)
tests/python/direct/test_with_id_model_indexer.py	5/5	Correctly consolidated dual skip conditions into single microarchitecture_is(10, 0) check

Sequence Diagram

sequenceDiagram
    participant Test as Test Runner
    participant Skip as Skip Condition Check
    participant Device as GPU Device
    participant Cutlass as Cutlass Scheduler
    participant Nvfuser as Nvfuser Operations
    
    Test->>Skip: Evaluate @pytest.mark.skipif
    Skip->>Device: Query compute capability
    Device-->>Skip: Returns (major, minor)
    
    alt Old Logic (is_pre_blackwell + microarchitecture_is_pre)
        Skip->>Skip: Check major >= 10 AND major < 12
        Note over Skip: Runs on 10.0, 10.3, 11.x
    else New Logic (microarchitecture_is)
        Skip->>Skip: Check major == 10 AND minor == 0
        Note over Skip: Only runs on 10.0
    else New Logic (is_blackwell - problematic)
        Skip->>Skip: Check 10.0, 10.3, 12.0, 12.1
        Note over Skip: Includes unsupported 12.0/12.1
    end
    
    alt Test NOT Skipped
        Test->>Nvfuser: Execute fusion (scaled_mm, cutlass ops)
        Nvfuser->>Cutlass: Request scheduling
        Cutlass->>Device: Verify compute capability
        alt Device is 10.0 or 10.3
            Cutlass-->>Nvfuser: Schedule operations
            Nvfuser-->>Test: Success
        else Device is 12.0 or 12.1
            Cutlass-->>Nvfuser: Error: Unsupported architecture
            Nvfuser-->>Test: FAIL
        end
    else Test Skipped
        Skip-->>Test: Skip test
    end

greptile-apps

_{5 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-01-05T16:28:11Z

tests/python/direct/test_narrow_precision.py

-@pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
+@pytest.mark.skipif(not is_blackwell(), reason="Only supported on blackwell.")


logic: Inconsistent architecture filtering: this test uses is_blackwell() (supports 10.0, 10.3, 12.0, 12.1) while other tests in this same file use microarchitecture_is(10, 0) (only supports 10.0). This creates different test coverage across Blackwell variants.

github-actions · 2026-01-05T16:30:29Z

Review updated until commit 835c7a8

Description

Replace compute capability checks with microarchitecture checks for test skipping
Add is_blackwell() utility function to identify Blackwell architectures (10.0, 10.3, 12.0, 12.1)
Update test decorators to skip unsupported architectures with clearer messaging
Remove deprecated is_pre_blackwell() and microarchitecture_is_pre() usage

Changes walkthrough

Relevant files

Tests

test_cutlass_nvfp4_gemm.py `Update NVFP4 GEMM test skipping logic` tests/python/direct/test_cutlass_nvfp4_gemm.py Remove global compute_cap check at module level Add `@pytest.mark.skipif` decorators to test functions using `microarchitecture_is(10, 0)` Import `microarchitecture_is` utility function Update skip reason to clarify Blackwell compute 12.0 is not supported	+13/-8
test_narrow_precision.py `Update narrow precision test skipping` tests/python/direct/test_narrow_precision.py Remove `is_pre_blackwell()` and `microarchitecture_is_pre(12)` skipif decorators Add `is_blackwell()` and `microarchitecture_is` imports Replace with `microarchitecture_is(10, 0)` checks for most tests Use `is_blackwell()` for one test requiring broader Blackwell support	+9/-15
test_with_id_model_indexer.py `Update model indexer test skipping` tests/python/direct/test_with_id_model_indexer.py Remove `is_pre_blackwell()` and `microarchitecture_is_pre(12)` skipif decorators Add `microarchitecture_is` import Add `@pytest.mark.skipif` decorator using `microarchitecture_is(10, 0)`	+3/-6

Enhancement

utils.py `Add Blackwell architecture detection utility` tests/python/direct_utils/utils.py Add `is_blackwell()` function to identify Blackwell architectures Support microarchitectures 10.0, 10.3, 12.0, and 12.1 Provide centralized utility for Blackwell architecture detection	+13/-0

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests
⚡ Recommended focus areas for review
Architecture Coverage Inconsistency The test decorators use `microarchitecture_is(10, 0)` which only allows compute capability 10.0, but the new `is_blackwell()` function in utils.py includes 10.3, 12.0, and 12.1 as Blackwell architectures. This suggests the tests might be too restrictive and not testing all intended Blackwell architectures. @pytest.mark.skipif( not microarchitecture_is(10, 0), reason="Does not support blackwell compute 12.0, other arches are not tested.", ) Inconsistent Architecture Checking There's inconsistent use of architecture checking functions - some tests use `microarchitecture_is(10, 0)` while others use `is_blackwell()`. This inconsistency should be reviewed to ensure all intended architectures are properly tested. @pytest.mark.skipif(not is_blackwell(), reason="Only supported on blackwell.") New Function Implementation The new `is_blackwell()` function should be validated to ensure it correctly identifies all intended Blackwell architectures (10.0, 10.3, 12.0, 12.1) and that this matches the intended test coverage. def is_blackwell(): return ( microarchitecture_is(10, 0) or microarchitecture_is(10, 3) or microarchitecture_is(12, 0) or microarchitecture_is(12, 1) )

Test failures

(Low, 1) Minor numerical mismatch in thunder instance_norm nvFuser CUDA test on A100

Test Name A100 Source

thunder.tests.test_ops.test_core_vs_torch_consistency_instance_norm_nvfuser_cuda_thunder.dtypes.float32 ❌

naoyam · 2026-01-05T18:51:05Z

The title suggests this has something to do with indexing. Can you clarify what it is?

liqiangxl · 2026-01-06T14:25:58Z

The title suggests this has something to do with indexing. Can you clarify what it is?

The title was not accurate, failure is not realated to indexing. Revised to skip unsupported arches in test test_with_id_model_indexer.py``

liqiangxl · 2026-01-06T14:31:03Z

@greptileai review this draft

greptile-apps

_{4 files reviewed, 4 comments}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-01-06T14:34:34Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


style: Inconsistent architecture filtering within same file: this test uses microarchitecture_is(10, 0) (only supports 10.0), but other tests in this file (lines 66, 173, 320) use is_blackwell() which supports all Blackwell variants (10.0, 10.3, 12.0, 12.1). Are the _scaled_mm tests specifically incompatible with 12.x architectures while quantization tests support them? Do scaled_mm operations have specific incompatibilities with 12.x compute capabilities that quantization operations don't have?

greptile-apps · 2026-01-06T14:34:35Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


style: Same inconsistency: microarchitecture_is(10, 0) here vs is_blackwell() at line 320 for similar grouped matmul tests

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

greptile-apps · 2026-01-06T14:34:35Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


style: Same pattern: microarchitecture_is(10, 0) here vs is_blackwell() elsewhere in file

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

greptile-apps · 2026-01-06T14:34:36Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


style: Same pattern: microarchitecture_is(10, 0) here vs is_blackwell() elsewhere in file

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

liqiangxl · 2026-01-06T14:47:04Z

!test

liqiangxl · 2026-01-06T15:46:48Z

Summary of failure reasons:
The first four tests are failed in the process of computing the reference output using PyTorch. The last test is due to nvFuser.

Test: CUDA_LAUNCH_BLOCKING=1 pytest tests/python/direct/test_cutlass_gemm.py -k "bfloat16-tokens_per_expert_neg_one and 115 and 1024" -vvvs
Err: Exception raised from run_group_mm at /opt/pytorch/nvfuser/cutlass/group_mm.cu:372
Reason: nvFuser grouped matmul related
Fix: already skipped in current main branch
Test: CUDA_LAUNCH_BLOCKING=1 pytest tests/python/direct/test_cutlass_nvfp4_gemm.py -k "gemm[shape and 256 and 64 and bfloat16" -vvvs
Err: Exception raised from runGemm at /opt/pytorch/nvfuser/cutlass/nvfp4_scaled_mm.cu:255
Reason: nvFuser scaled matmul related
Test CUDA_LAUNCH_BLOCKING=1 pytest tests/python/direct/test_cutlass_nvfp4_gemm.py -k "grouped_mm and bfloat"
Err: Exception raised from run_nvfp4_scaled_group_mm at /opt/pytorch/nvfuser/cutlass/nvfp4_scaled_group_mm.cu:518
Reason: nvFuser scaled-grouped matmul related
Test CUDA_LAUNCH_BLOCKING=1 pytest tests/python/direct/test_narrow_precision.py -k "scaled_mm" -vvvs
Reason: nvFuser scaled matmul related
Test NVFUSER_DISABLE=parallel_compile pytest tests/python/direct/test_with_id_model_indexer.py -k "eager" -vvvs
Err: INTERNAL ASSERT FAILED at /opt/pytorch/nvfuser/csrc/runtime/executor_utils.cpp:559
Reason: seems nvFuser related.

liqiangxl · 2026-01-13T13:51:29Z

!test

greptile-apps

_{6 files reviewed, 6 comments}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-01-13T13:57:31Z

tests/python/direct_utils/utils.py

+def is_blackwell():
+    return (
+        microarchitecture_is(10, 0)
+        or microarchitecture_is(10, 3)
+        or microarchitecture_is(12, 0)
+        or microarchitecture_is(12, 1)
+    )


The is_blackwell() function incorrectly includes compute capability 12.0 and 12.1 as "Blackwell" architectures. However, the cutlass scheduler (csrc/scheduler/cutlass.cpp lines 75-84) explicitly only supports compute capabilities 10.0 and 10.3:

if (device_prop->major != 10 || !(device_prop->minor == 0 || device_prop->minor == 3)) { // error: "Cutlass scheduler only supports GB200 and GB300 (cc 10.0 or 10.3)" }

Compute capabilities 12.0 (RTX PRO 6000 and RTX 50XX) and 12.1 (DGX Spark) are NOT Blackwell architectures and are NOT supported by the cutlass scheduler.

This function should only return true for 10.0 and 10.3:

Suggested change

def is_blackwell():

return (

microarchitecture_is(10, 0)

or microarchitecture_is(10, 3)

or microarchitecture_is(12, 0)

or microarchitecture_is(12, 1)

)

def is_blackwell():

return (

microarchitecture_is(10, 0)

or microarchitecture_is(10, 3)

)

Alternatively, if 12.0 and 12.1 are legitimately Blackwell variants, the function should be renamed to clarify it's broader than what cutlass supports, and tests should use more specific checks.

greptile-apps · 2026-01-13T13:57:32Z

tests/python/direct/test_narrow_precision.py

-@pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
+@pytest.mark.skipif(not is_blackwell(), reason="Only supported on blackwell.")


Inconsistent architecture check: this test uses is_blackwell() which includes compute capabilities 10.0, 10.3, 12.0, and 12.1. However:

Other similar tests in this file that use fd.ops.scaled_mm (like test_scaled_mm at line 245) use microarchitecture_is(10, 0) - only 10.0

The cutlass scheduler only supports 10.0 and 10.3 (not 12.0 or 12.1)

This test uses fd.ops.scaled_mm and fd.ops.nv_block_quantize, which may rely on cutlass scheduling. If so, it should use the same skip condition as other tests:

Suggested change

@pytest.mark.skipif(not is_blackwell(), reason="Only supported on blackwell.")

@pytest.mark.skipif(not microarchitecture_is(10, 0), reason="Only supported on blackwell.")

If this test is genuinely intended to support all Blackwell variants (10.0, 10.3, 12.0, 12.1) because scaled_mm uses a different scheduler (nvmmh), then the is_blackwell() function needs to be fixed to exclude 12.0 and 12.1, OR this needs documentation explaining why it's different.

greptile-apps · 2026-01-13T13:57:32Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


The skip condition logic has changed significantly. The old conditions were:

is_pre_blackwell() → skip if major < 10

not microarchitecture_is_pre(12) → skip if major >= 12

This ran tests on architectures where: 10 <= major < 12 (i.e., 10.0, 10.3, and any 11.x devices)

The new condition not microarchitecture_is(10, 0) runs tests ONLY on exactly 10.0.

If the intent is to support 10.3 (B300/GB300) in addition to 10.0, the condition should be:

Suggested change

@pytest.mark.skipif(

is_pre_blackwell(), reason="Only supported on blackwell and newer devices."

)

@pytest.mark.skipif(

not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"

not microarchitecture_is(10, 0),

reason="Does not support blackwell compute 12.0, other arches are not tested.",

)

@pytest.mark.skipif(

not (microarchitecture_is(10, 0) or microarchitecture_is(10, 3)),

reason="Does not support blackwell compute 12.0, other arches are not tested.",

)

However, if 10.3 is deliberately excluded because it hasn't been tested, the current change is correct but note that this is more restrictive than the original logic.

greptile-apps · 2026-01-13T13:57:34Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


Same issue as test_scaled_mm: the skip condition logic has changed from supporting 10.x and 11.x architectures to only supporting exactly 10.0.

Old: is_pre_blackwell() AND not microarchitecture_is_pre(12) → run on 10 <= major < 12
New: not microarchitecture_is(10, 0) → run only on major==10 and minor==0

This test uses fd.ops.cutlass_nvfp4_grouped_mm which is explicitly a cutlass operation. The cutlass scheduler supports both 10.0 and 10.3 according to csrc/scheduler/cutlass.cpp. Consider updating to support both tested architectures if 10.3 should be included.

greptile-apps · 2026-01-13T13:57:35Z

tests/python/direct/test_narrow_precision.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


Same issue: the skip condition has become more restrictive, changing from supporting 10.x and 11.x to only 10.0. This test also uses fd.ops.cutlass_nvfp4_grouped_mm which should support both 10.0 and 10.3 based on the cutlass scheduler implementation.

greptile-apps · 2026-01-13T13:57:36Z

tests/python/direct/test_with_id_model_indexer.py

 @pytest.mark.skipif(
-    is_pre_blackwell(), reason="Only supported on blackwell and newer devices."
-)
-@pytest.mark.skipif(
-    not microarchitecture_is_pre(12), reason="Does not support blackwell compute 12.0"
+    not microarchitecture_is(10, 0),
+    reason="Does not support blackwell compute 12.0, other arches are not tested.",
 )


The skip condition has changed from supporting 10.x and 11.x architectures (via is_pre_blackwell() AND not microarchitecture_is_pre(12)) to only supporting exactly 10.0. This is consistent with the changes in test_narrow_precision.py but represents a more restrictive test coverage than before. Verify this is intentional.

liqiangxl added 4 commits January 5, 2026 07:15

fix test_grouped_mm

4b7d91b

skip nvfp4

235f161

fix narrow precision tests

40f0ef7

fix test_with_id_model_indexer

f9fdc32

greptile-apps bot reviewed Jan 5, 2026

View reviewed changes

liqiangxl changed the title ~~fix test with id model indexer~~ skip unsupported arches in test test_with_id_model_indexer.py Jan 6, 2026

liqiangxl marked this pull request as draft January 6, 2026 14:30

Merge branch 'main' into llu/fix_test_with_id_model_indexer

55b5572

greptile-apps bot reviewed Jan 6, 2026

View reviewed changes

liqiangxl changed the title ~~skip unsupported arches in test test_with_id_model_indexer.py~~ skip unsupported arches in python tests Jan 6, 2026

liqiangxl and others added 3 commits January 12, 2026 15:50

Merge branch 'main' into llu/fix_test_with_id_model_indexer

bd58ae9

skip scaledMM groupedMM quantLayout

e3e0b6a

Merge branch 'main' into llu/fix_test_with_id_model_indexer

835c7a8

liqiangxl marked this pull request as ready for review January 13, 2026 13:51

greptile-apps bot reviewed Jan 13, 2026

View reviewed changes

	@pytest.mark.skipif(not is_blackwell(), reason="Only supported on blackwell.")
	@pytest.mark.skipif(not microarchitecture_is(10, 0), reason="Only supported on blackwell.")

skip unsupported arches in python tests #5760

Are you sure you want to change the base?

skip unsupported arches in python tests #5760

Conversation

liqiangxl commented Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liqiangxl commented Jan 5, 2026

Uh oh!

greptile-apps bot commented Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Overview

Greptile Summary

Overview

Key Changes

Critical Issues Found

1. Incorrect Architecture Definition in is_blackwell()

2. Inconsistent Skip Conditions in test_narrow_precision.py

3. More Restrictive Than Original Logic

Recommendations

Confidence Score: 2/5

Important Files Changed

Sequence Diagram

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions bot commented Jan 5, 2026 • edited by xwang233 Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Changes walkthrough

PR Reviewer Guide

Test failures

Uh oh!

naoyam commented Jan 5, 2026

Uh oh!

liqiangxl commented Jan 6, 2026

Uh oh!

liqiangxl commented Jan 6, 2026

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 6, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 6, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 6, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 6, 2026

Choose a reason for hiding this comment

Uh oh!

liqiangxl commented Jan 6, 2026

Uh oh!

liqiangxl commented Jan 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liqiangxl commented Jan 13, 2026

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

liqiangxl commented Jan 5, 2026 •

edited

Loading

greptile-apps bot commented Jan 5, 2026 •

edited

Loading

1. Incorrect Architecture Definition in `is_blackwell()`

2. Inconsistent Skip Conditions in `test_narrow_precision.py`

github-actions bot commented Jan 5, 2026 •

edited by xwang233

Loading

liqiangxl commented Jan 6, 2026 •

edited

Loading