[MetaXGPU] apply cuda tests to maca, mark with xfail if failed by JoieAli · Pull Request #40 · MetaX-MACA/mcTVM

JoieAli · 2026-07-10T02:42:36Z

No description provided.

gemini-code-assist

Code Review

This pull request transitions various GPU-related tests and support modules from CUDA/NVIDIA configurations to MACA (Meta Accelerator Card Architecture), renaming targets, devices, and imports, and marking unsupported features as expected failures. The code review feedback correctly identifies several remaining NVIDIA-specific configurations in the migrated tests, such as the use of tvm.support.nvcc.get_target_compute_version() instead of the MACA-equivalent mxcc helper, and the inclusion of the CUDA-specific "arch" attribute in target configurations, which should be removed or replaced with "mcpu" for "maca" targets.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

gemini-code-assist · 2026-07-10T02:46:05Z

+)
 def test_inject_ptx_intrin():
    f = vector_add
    arch = tvm.support.nvcc.get_target_compute_version()


Since this test is skipped if not env.has_maca(), when it runs, it will execute on a MACA device. However, tvm.support.nvcc.get_target_compute_version() is NVIDIA-specific and will fail or return incorrect results on a MACA device.

Please update this to use tvm.support.mxcc.get_target_compute_version() instead.

Suggested change

arch = tvm.support.nvcc.get_target_compute_version()

from tvm.support import mxcc

arch = mxcc.get_target_compute_version()

gemini-code-assist · 2026-07-10T02:46:06Z

        kind="cuda",
        mod=mod,
-        target=Target({"kind": "cuda", "arch": "sm_70"}),
+        target=Target({"kind": "maca", "arch": "sm_70"}),


The "arch" attribute (e.g., "sm_70") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu" (e.g., "xcore700"), and passing "arch" will likely result in target creation errors or be ignored.

Since the specific architecture is not strictly required for this test, please remove "arch": "sm_70" from the target configuration, similar to the changes made in test_transform_realize_vdevice.py and test_tvmscript_parser.py.

Suggested change

target=Target({"kind": "maca", "arch": "sm_70"}),

target=Target({"kind": "maca"}),

gemini-code-assist · 2026-07-10T02:46:06Z

        kind="cuda",
        mod=mod,
-        target=tvm.target.Target({"kind": "cuda", "arch": "sm_70"}),
+        target=tvm.target.Target({"kind": "maca", "arch": "sm_70"}),


The "arch" attribute (e.g., "sm_70") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_70" from the target configuration here and in other occurrences within this file.

Suggested change

target=tvm.target.Target({"kind": "maca", "arch": "sm_70"}),

target=tvm.target.Target({"kind": "maca"}),

gemini-code-assist · 2026-07-10T02:46:06Z

    config = {"mbar": Var("mbar", "handle"), "remote_cta_id": IntImm("int32", 1)}
    op_call = CopyAsync(BufferRegion(dst_buf, ranges), BufferRegion(src_buf, ranges), config=config)
-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_90a"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_90a"})


The "arch" attribute (e.g., "sm_90a") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_90a" from the target configuration here and in other occurrences within this file.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_90a"})

target = tvm.target.Target({"kind": "maca"})

gemini-code-assist · 2026-07-10T02:46:06Z

    op_call = CopyAsync(dst_br, src_br, config=config)

-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_90a"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_90a"})


The "arch" attribute (e.g., "sm_90a") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_90a" from the target configuration.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_90a"})

target = tvm.target.Target({"kind": "maca"})

gemini-code-assist · 2026-07-10T02:46:06Z

    # when no GPU is detected, which nvcc 13+ rejects. Codegen happens before
    # nvcc; if the whole tvm.compile pipeline fails, we never see the source.
-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_90"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_90"})


The "arch" attribute (e.g., "sm_90") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_90" from the target configuration.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_90"})

target = tvm.target.Target({"kind": "maca"})

gemini-code-assist · 2026-07-10T02:46:06Z

    """Full pipeline (UnrollLoop + CUDA codegen) emits one mma per (Mt, Nt, Kt)
    tile; K-tiles accumulate in place, so D is cleared once per output tile."""
-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_80"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_80"})


The "arch" attribute (e.g., "sm_80") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_80" from the target configuration.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_80"})

target = tvm.target.Target({"kind": "maca"})

gemini-code-assist · 2026-07-10T02:46:06Z

        Tx.copy(A[tx], ra)

-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_100a"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_100a"})


The "arch" attribute (e.g., "sm_100a") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_100a" from the target configuration here and in other occurrences within this file.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_100a"})

target = tvm.target.Target({"kind": "maca"})

gemini-code-assist · 2026-07-10T02:46:06Z

        Tx.copy(D[tx], rd)

-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_100a"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_100a"})


The "arch" attribute (e.g., "sm_100a") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_100a" from the target configuration.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_100a"})

target = tvm.target.Target({"kind": "maca"})

gemini-code-assist · 2026-07-10T02:46:06Z


        # Use sm_100a target for packed add sum dispatch
-    target = tvm.target.Target({"kind": "cuda", "arch": "sm_100a"})
+    target = tvm.target.Target({"kind": "maca", "arch": "sm_100a"})


The "arch" attribute (e.g., "sm_100a") is specific to the "cuda" target kind. For the "maca" target kind, the appropriate attribute is "mcpu", and passing "arch" will likely result in target creation errors or be ignored.

Please remove "arch": "sm_100a" from the target configuration.

Suggested change

target = tvm.target.Target({"kind": "maca", "arch": "sm_100a"})

target = tvm.target.Target({"kind": "maca"})

Five-HZ

LGTM

gemini-code-assist Bot reviewed Jul 10, 2026

View reviewed changes

JoieAli force-pushed the mcTVM_apply_cuda_cases branch from 039d9e1 to 678c078 Compare July 10, 2026 03:54

[MetaXGPU] apply cuda tests to maca, mark with xfail if failed

e93b7da

JoieAli force-pushed the mcTVM_apply_cuda_cases branch from 678c078 to e93b7da Compare July 10, 2026 05:55

JoieAli added 2 commits July 10, 2026 09:34

[MetaXGPU][CI] delete PYTEST_ADDOPTS

00a3601

[MetaXGPU][CI] Add report after running maca test

80e15f6

Five-HZ approved these changes Jul 13, 2026

View reviewed changes

Five-HZ merged commit 726b6cf into MetaX-MACA:dev Jul 13, 2026
5 checks passed

	arch = tvm.support.nvcc.get_target_compute_version()
	from tvm.support import mxcc
	arch = mxcc.get_target_compute_version()

	target=Target({"kind": "maca", "arch": "sm_70"}),
	target=Target({"kind": "maca"}),

	target=tvm.target.Target({"kind": "maca", "arch": "sm_70"}),
	target=tvm.target.Target({"kind": "maca"}),

	target = tvm.target.Target({"kind": "maca", "arch": "sm_90a"})
	target = tvm.target.Target({"kind": "maca"})

Uh oh!

Conversation

JoieAli commented Jul 10, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 10, 2026

Choose a reason for hiding this comment

Uh oh!

Five-HZ left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants