Print TensorView for debugging

[wujingyue]

Otherwise I couldn't even locate the Expr in a printed fusion.

[wujingyue]

!test

[github-actions]

Description

• Enhanced error messages in broadcast validation by including TensorView string representations

• Replaced generic mesh rank validation messages with detailed debugging information

• Added input_tv->toString() and output_tv->toString() to error messages for better debugging

Changes walkthrough

	Relevant files
Enhancement	lower_to_communication.cpp Enhanced error messages with TensorView debugging                csrc/host_ir/lower_to_communication.cpp Modified NVF_ERROR_EQ calls for sender_mesh and receiver_mesh rank validation Replaced generic error messages with detailed TensorView string representations Added toString() calls to input_tv and output_tv for debugging purposes +2/-10

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Error Message Improvement The PR successfully replaces generic error messages with more informative ones that include TensorView string representations. This change should help developers debug broadcast operations by providing tensor view details in error messages. The modification is safe and maintains the same error checking logic while improving debuggability. NVF_ERROR_EQ(sender_mesh.rank(), 1, "sender: ", input_tv->toString()); NVF_ERROR_EQ(receiver_mesh.rank(), 1, "receiver: ", output_tv->toString());

Test failures (partial, pipeline still running)

• (Medium, 1) Thunder nvFuser scalar mismatch in nanogpt autograd test (test_networks)

  Test Name	A100	Source
  thunder.tests.test_networks.test_nanogpt_complete_autograd_nvfuser_cuda_thunder.dtypes.float32	❌

[greptile-apps]

Greptile Overview

Greptile Summary

This PR adjusts the error messages in lowerToBroadcast to print TensorView strings when validating that sender/receiver device meshes are 1D, making it easier to locate the corresponding TV/Expr when inspecting a printed fusion.

The functional behavior of broadcast lowering remains the same; only the diagnostics emitted on rank mismatch are changed.

Confidence Score: 4/5

• This PR is likely safe to merge; it only changes diagnostic strings in an error path.
• Change is localized to lowerToBroadcast and does not alter the lowering logic; main concern is a potential loss of useful debug context by no longer printing the offending DeviceMesh values.
• csrc/host_ir/lower_to_communication.cpp

Important Files Changed

Filename	Overview
csrc/host_ir/lower_to_communication.cpp	Tweaks Broadcast rank-check error messages to print TensorView strings instead of the device meshes; behavior unchanged but debug context may be less informative.

Sequence Diagram

sequenceDiagram
  participant Caller as convertSingleOpToCommunication
  participant Info as getCommunicationInfo
  participant Bcast as lowerToBroadcast
  participant Mesh as DeviceMesh

  Caller->>Info: compute CommunicationInfo(type)
  Info-->>Caller: type = Broadcast
  Caller->>Bcast: lowerToBroadcast(input_tv, output_tv)
  Bcast->>Mesh: sender_mesh = input_tv.getDeviceMesh()
  Bcast->>Mesh: receiver_mesh = output_tv.getDeviceMesh()
  Bcast->>Bcast: NVF_ERROR_EQ(sender_mesh.rank(), 1, ...)
  Bcast->>Bcast: NVF_ERROR_EQ(receiver_mesh.rank(), 1, ...)
  Bcast->>Mesh: root = sender_mesh.at(0)
  Bcast->>Bcast: team = receiver_mesh.vector()
  Bcast-->>Caller: comms.push_back(Communication::Broadcast)

Loading

[greptile-apps]

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

[P2] Error message regression: the previous message included the actual DeviceMesh ("Given ", sender_mesh/receiver_mesh). The new message prints TensorView::toString() but no longer prints the offending mesh, which can make debugging rank mismatches harder (you can’t see whether rank() is wrong because the mesh is unexpected vs the TV being different).

Consider keeping the original context and appending the TV string, e.g. include sender_mesh/receiver_mesh in the message as well.