feat: refine mxfp8 quant kernel

[RuibinCheung]

• Implement MXFP8 hip kernel.
• Remove MXFP4 and MXFP8 triton kernel.
• Fix bug in mxfp4 quant kernel when enable 2d block.

[Copilot]

Pull request overview

This PR migrates MXFP8 quantization from Triton to a new HIP/CUDA kernel exposed via the PyTorch C++ extension, while removing the previous Triton MXFP4/MXFP8 kernels and adjusting FP8 GEMM dtype selection logic.

Changes:

• Add C++/HIP implementations + bindings for quantize_mxfp8 and quantize_mxfp8_dual (with shuffle support) and route Python MXFP8 quantization to these ops.
• Remove Triton MXFP4/MXFP8 kernel source files.
• Update tests and FP8 GEMM helpers to use centralized FP8 dtype selection and MX block-size constants.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/pytorch/ops/test_quantization.py	Updates MXFP8 test coverage and adds a shuffle-scale test.
primus_turbo/triton/quantization/quantization_mxfp8.py	Removes Triton MXFP8 quant/dequant kernels.
primus_turbo/triton/quantization/quantization_mxfp4.py	Removes Triton MXFP4 quant/dequant kernels.
primus_turbo/pytorch/ops/grouped_gemm_fp8.py	Centralizes FP8 dtype selection (adds HYBRID support).
primus_turbo/pytorch/ops/gemm_fp8.py	Centralizes FP8 dtype selection (adds HYBRID support).
primus_turbo/pytorch/kernels/quantization/quantization_impl.py	Switches MXFP8 quantization to C++ ops; still references Triton dequant kernels.
primus_turbo/pytorch/core/low_precision.py	Introduces MXFP4/MXFP8 block-size constants used by configs/tests.
csrc/pytorch/quantization/quantization_meta.cpp	Adds meta implementations for MXFP8 quant ops.
csrc/pytorch/quantization/quantization.cpp	Implements MXFP8 quant ops and dispatch into new kernels.
csrc/pytorch/extensions.h	Declares new MXFP8 quant APIs.
csrc/pytorch/bindings_pytorch.cpp	Registers new MXFP8 ops with CUDA/Meta impls.
csrc/kernels/quantization/quantization_mxfp8.cu	Adds the new MXFP8 HIP/CUDA quantization kernels (single + dual).
csrc/kernels/quantization/quantization_mxfp4.cu	Fixes/cleans up shuffle + reduction utilities usage.
csrc/include/primus_turbo/shuffle.h	Adds MXFP8 shuffle layout constants.
csrc/include/primus_turbo/quantization.h	Adds MXFP8 constants/types + function declarations.
csrc/include/primus_turbo/device/utils.cuh	Consolidates low-level bitcast + fp16/bf16 unpack helpers.
csrc/include/primus_turbo/device/shuffle.cuh	Generalizes shuffled-index computation for FP4/FP8.
csrc/include/primus_turbo/device/reduce.cuh	Adds AMD wavefront DPP-based max reductions used by kernels.

---

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

[Copilot]

The NOTE comment refers to the “fp4 kernel” and CDNA4 ISA, but this function is quantize_mxfp8_impl and uses the MXFP8 backend. Please update the comment to accurately describe the MXFP8 kernel/support check to avoid confusion during maintenance.

Suggested change

	# NOTE: quantize fp4 kernel use the ISA which only available on cdna4.
	# NOTE: MXFP8 quantization kernels rely on ISA features only available on CDNA4;
	# check_mxfp8_support() verifies that this hardware support is present.

[Copilot]

In the MXFP8 shuffle test, quantize_fp8_with_trans(..., block_size=MXFP4_BLOCK_SIZE) is confusing/misleading since this is an FP8 path. Please use MXFP8_BLOCK_SIZE here to match intent and avoid accidental divergence if FP4/FP8 block sizes ever differ.

Suggested change

	block_size=MXFP4_BLOCK_SIZE,
	block_size=MXFP8_BLOCK_SIZE,

[Copilot]

primus_turbo.triton.quantization.quantization_mxfp4 / quantization_mxfp8 no longer exist (the directory now only contains quant_blockwise.py), so these imports will raise ModuleNotFoundError at import time and break MXFP4/MXFP8 dequantization paths that still reference the Triton kernels. Either restore the Triton dequantize kernels/modules or migrate dequantize_mxfp4_impl/dequantize_mxfp8_impl to the new C++/HIP backend and remove these imports.