[DRAFT] Update rocm/flashinfer to v0.5.3 upstream tag

[diptorupd]

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

This PR updates FlashInfer to the v0.5.3 upstream tag, transitioning from PyTorch-specific bindings to a framework-agnostic TVM FFI approach. The changes involve replacing PyTorch tensor types with TensorView abstractions and updating error handling mechanisms throughout the codebase.

Changes:

• Migration from PyTorch (at::Tensor) to TVM FFI framework (TensorView) for tensor operations
• Replacement of PyTorch error handling (TORCH_CHECK) with TVM FFI checks (TVM_FFI_ICHECK)
• Updates to benchmarking utilities replacing Triton's do_bench with custom bench_gpu_time
• Infrastructure improvements including Docker configuration updates and GitHub workflow enhancements
• Code quality improvements with Python linting fixes

Reviewed changes

Copilot reviewed 123 out of 849 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
csrc/cascade.cu	Migrated from PyTorch tensors to TensorView with TVM FFI error handling
csrc/bmm_fp8.cu	Updated tensor types and error checking for FP8 GEMM operations
csrc/blackwell_fmha_plan.cu	New file for Blackwell FMHA planning functionality
csrc/batch_prefill_*.cu	Refactored prefill operations with TVM FFI bindings
csrc/batch_decode*.cu	Updated decode operations with new tensor abstractions
csrc/batch_mla*.cu	MLA operation updates for TVM FFI compatibility
benchmarks/*.py	Replaced Triton benchmarking with custom utilities
ci/scripts/jenkins/git_utils.py	Python code quality improvements
build_backend.py	New build system backend implementation
.github/workflows/*.yml	CI/CD workflow updates and new automation

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The type annotation Dict[Any, Any] is too permissive. Headers should be Dict[str, str] since HTTP headers are string key-value pairs.

Suggested change

	headers: Dict[Any, Any] = {}
	headers: Dict[str, str] = {}

[Copilot]

The error message construction concatenates strings with newlines. Consider using a more structured format or logging the original exception details separately for better error tracking.

Suggested change

	raise RuntimeError(f"Error response: {msg}\n{error_data}") from e
	logging.error(
	"HTTPError during request",
	extra={"error_message": msg, "error_data": error_data},
	)
	raise RuntimeError(
	f"Error response: {msg}; see logs for error details"
	) from e

[Copilot]

Missing colon after 'fmha' in error message for consistency with other error messages in the codebase (e.g., line 125 in batch_mla_sm90_run.cu has 'error:').

Suggested change

	<< "Failed to plan blackwell fmha" << cudaGetErrorString(status);
	<< "Failed to plan blackwell fmha: " << cudaGetErrorString(status);

[Copilot]

Function name init_seed_sampling_from_softmax_logits is misleading as it suggests the function initializes a seed, but it actually performs sampling. Consider renaming to sampling_from_softmax_logits for clarity.