Unconditionally return 1 for ParallelType::Stream for unshardedSizes

[zasdfgbnm]

No description provided.

[zasdfgbnm]

!test

[github-actions]

Description

• Simplified ParallelType::Stream handling in unshardedSizes function

• Removed conditional logic for stream parallelization dimension checking

• Eliminated error checking for non-constant stream extent

• Unconditionally returns 1 for all stream parallelized dimensions

Changes walkthrough

Relevant files

Bug fix

execution_utils.cpp Simplify stream parallelization handling in unshardedSizes csrc/multidevice/execution_utils.cpp Removed TODO comment about MultiDeviceExecutor hack for stream parallelization Eliminated conditional logic checking if sharded_id is in logical domain Removed error checking for non-constant stream extent evaluation Simplified to unconditionally return 1 for ParallelType::Stream +1/-23

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review

Removed Error Checking The PR removes important error checking that validated DIDs/Stream extent is constant. This error checking was ensuring that non-constant extents would be caught early rather than potentially causing runtime issues. The removal of this validation could lead to silent failures or incorrect behavior when non-constant extents are encountered. return 1; Behavioral Change Risk The change removes the conditional logic that only returned 1 for logical domains while evaluating extent for other domains. The new unconditional return of 1 for all ParallelType::Stream cases could be incorrect for non-logical domains and may break existing functionality that depends on the extent evaluation. if (parallel_type == ParallelType::Stream) { return 1; } Missing Documentation The PR removes a detailed TODO comment that explained the reasoning behind the original implementation. This context is valuable for understanding why certain decisions were made and what future work is planned. The removal of this documentation makes it harder for future maintainers to understand the design rationale. if (parallel_type == ParallelType::Stream) { return 1; }

Test failures

• (High, 95) CUDA driver/runtime mismatch causing init-time failures in nvFuser matmul & top-k test suites on dlcluster_h100

  Test Name	H100	Source
  Ampere/MmaTest.SingleTile/Ampere_16_8_16__bfloat	❌	Link
  ArgsortParameterizedWithBlockAndBatch.SharedMemoryRequirement/2048_1_1_0	❌	Link
  BlockSizeAndItemsPerThread/ArgSortComprehensiveTest.ComprehensiveValidation/BlockSize32_ItemsPerThread4	❌	Link
  ClusterReductionTest.SimpleFusionNotAllReduce/cluster_15_dtype_double	❌	Link
  ClusterReductionTest.SimpleFusionNotAllReduce/cluster_4_dtype_double	❌	Link
  CutlassExecutorTest.Nvfp4Matmul_BiasEpilogue	❌	Link
  General/HopperPlusMatmulSchedulerTest.FusedMultiplySum/KK_512_256_128_MmaMacro_m64_n128_k16_splitk_2	❌	Link
  General/HopperPlusMatmulSchedulerTest.FusedMultiplySum/MK_512_256_128_MmaMacro_m128_n128_k16_tma_store	❌	Link
  General/HopperPlusMatmulSchedulerTest.FusedMultiplySumBiasNeg/MN_512_256_128_MmaMacro_m64_n128_k16_tma_store_splitk_2	❌	Link
  GreedySchedulerTest.ScanNonLocalOutput	❌	Link
  ... with 85 more test failures omitted. Check internal logs.

• (High, 1) Outdated NVIDIA driver on dlcluster_h100 causing CUDA initialization failure in RNG tests

  Test Name	H100	Source
  RNGTest.BroadcastingRNG	❌	Link

• (Medium, 12) nvFuser internal assert on non-divisible split (tensor_metadata.cpp) in test_stream.test_two_matmuls_inlinable and multidevice.test_overlap suites

  Test Name	A100	A100 (dist.)	GB200	GB200 (dist.)	H100	H100 (dist.)	Source
  tests.python.direct.test_stream.test_two_matmuls_inlinable[nvfuser_direct_test=eager]	❌		❌		❌
  tests.python.direct.test_stream.test_two_matmuls_inlinable[nvfuser_direct_test=lru_cache]	❌		❌		❌
  tests.python.multidevice.test_overlap.test_row_parallel_linear_forward	❌	❌	❌	❌	❌	❌

• (Low, 1) Small numerical mismatch in nvFuser reduction tutorial test (tests.python.direct.test_tutorial)

  Test Name	GB200	Source
  tests.python.direct.test_tutorial.test_tutorial_reduction	❌