[mxfp8 moe training] add CUDA kernel for per-group conversion of scale factors to blocked layout

stack-info: PR: #3504, branch: danielvegamyhre/stack/86

Author: danielvegamyhre

benchmarks/prototype/moe_training/mxfp8/bench_triton_mx_block_rearrange_2d_K_groups.py

@@ -0,0 +1,224 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD 3-Clause license found in the
+# LICENSE file in the root directory of this source tree.
+
+import itertools
+from dataclasses import dataclass
+from typing import List
+
+import torch
+from tabulate import tabulate
+from tqdm import tqdm
+
+from benchmarks.utils import benchmark_cuda_function_in_microseconds
+from torchao.prototype.moe_training.kernels.mxfp8.quant import (
+    mx_block_rearrange_2d_K_groups_cuda,
+    torch_to_blocked_2d_K_groups,
+    triton_mx_block_rearrange_2d_K_groups,
+)
+from torchao.prototype.moe_training.utils import generate_jagged_offs
+
+device = torch.device("cuda")
+
+# Needed since changing args to function causes recompiles
+torch._dynamo.config.cache_size_limit = 1000
+
+
+@dataclass(frozen=True)
+class ExperimentConfig:
+    input_shape: tuple[int]
+    num_groups: int
+    version: str  # "naive" or "parallel"
+
+
+@dataclass(frozen=True)
+class ExperimentResult:
+    time_us: float
+    mem_bw_gbps: float
+
+
+@dataclass(frozen=True)
+class Experiment:
+    config: ExperimentConfig
+    result: ExperimentResult
+
+
+def get_configs() -> List[ExperimentConfig]:
+    # Llama4 and DSV3 671b shapes. Input activations are scaled along the total_M dim, which contains all the token groups.
+    block_size = 32
+    input_shapes = [
+        (8192, 16384 // block_size),
+        (8192, 32768 // block_size),
+        (8192, 65536 // block_size),
+        (8192, 131072 // block_size),
+        (8192, 1048576 // block_size),
+        (5120, 16384 // block_size),
+        (5120, 32768 // block_size),
+        (5120, 65536 // block_size),
+        (5120, 131072 // block_size),
+        (5120, 1048576 // block_size),
+        (7168, 16384 // block_size),
+        (7168, 32768 // block_size),
+        (7168, 65536 // block_size),
+        (7168, 131072 // block_size),
+        (7168, 1048576 // block_size),
+        (2048, 16384 // block_size),
+        (2048, 32768 // block_size),
+        (2048, 65536 // block_size),
+        (2048, 131072 // block_size),
+        (2048, 1048576 // block_size),
+    ]
+    num_groups = [8]
+    versions = [
+        "torch",
+        "triton",
+        "cuda_pipelined_64_chunks4",
+    ]
+
+    configs = []
+    for shape, groups, version in itertools.product(
+        input_shapes,
+        num_groups,
+        versions,
+    ):
+        configs.append(
+            ExperimentConfig(
+                input_shape=shape,
+                num_groups=groups,
+                version=version,
+            )
+        )
+    return configs
+
+
+def run_experiment(config: ExperimentConfig) -> ExperimentResult:
+    input_shape, num_groups, version = (
+        config.input_shape,
+        config.num_groups,
+        config.version,
+    )
+    input_tensor = torch.randint(
+        low=0,
+        high=256,
+        size=input_shape,
+        dtype=torch.uint8,
+        device=device,
+    )
+
+    M, Kg = input_shape
+    block_size = 32
+    input_group_offsets = generate_jagged_offs(num_groups, Kg, multiple_of=block_size)
+
+    # Select which kernel to benchmark based on version
+    if version == "torch":
+        kernel_fn = torch_to_blocked_2d_K_groups
+        kernel_input = input_tensor
+    elif version == "triton":
+        kernel_fn = triton_mx_block_rearrange_2d_K_groups
+        # Triton uses row-major input
+        kernel_input = input_tensor
+    elif version == "cuda_pipelined_64_chunks4":
+        kernel_fn = lambda t, o: mx_block_rearrange_2d_K_groups_cuda(
+            t,
+            o,
+        )
+        kernel_input = input_tensor.view(torch.float8_e8m0fnu)
+    else:
+        raise ValueError(f"Unknown version: {version}")
+
+    # Run kernel to get output shape
+    outputs = kernel_fn(
+        kernel_input,
+        input_group_offsets,
+    )
+    if isinstance(outputs, tuple):  # torch returns a tuple with extra metadata
+        out_scales, _ = outputs
+    else:
+        out_scales = outputs
+
+    # Benchmark the kernel
+    time_us = benchmark_cuda_function_in_microseconds(
+        kernel_fn,
+        kernel_input,
+        input_group_offsets,
+    )
+
+    # Calculate memory bandwidth
+    bytes_per_input_el = torch.finfo(torch.float8_e8m0fnu).bits / 8
+    bytes_per_output_el = torch.finfo(torch.float8_e4m3fn).bits / 8
+
+    read_bytes = input_tensor.numel() * bytes_per_input_el
+    write_bytes = out_scales.numel() * bytes_per_output_el
+
+    mem_bw_gbps = ((read_bytes + write_bytes) / 1e9) / (time_us / 1e6)
+
+    return ExperimentResult(
+        time_us=time_us,
+        mem_bw_gbps=mem_bw_gbps,
+    )
+
+
+def print_results(experiments: List[Experiment]):
+    # Group experiments by input shape
+    shapes_dict = {}
+    for exp in experiments:
+        shape_key = exp.config.input_shape
+        if shape_key not in shapes_dict:
+            shapes_dict[shape_key] = {}
+        shapes_dict[shape_key][exp.config.version] = exp.result
+
+    headers = [
+        "kernel_version",
+        "scale_shape",
+        "time_us",
+        "mem_bw_gbps",
+        "fastest_version",
+        "speedup_vs_torch",
+    ]
+
+    rows = []
+    for shape, versions in shapes_dict.items():
+        # Find fastest version for this shape
+        fastest_version = min(versions.items(), key=lambda x: x[1].time_us)[0]
+
+        # Get torch baseline time for speedup calculation
+        torch_time_us = versions.get("torch").time_us if "torch" in versions else None
+
+        # Add rows for each version
+        for version, result in versions.items():
+            # Calculate speedup vs triton
+            speedup_str = ""
+            if version != "torch":
+                speedup = torch_time_us / result.time_us
+                speedup_str = f"{speedup:.2f}x"
+
+            rows.append(
+                [
+                    version,
+                    f"({shape[0]}, {shape[1]})",
+                    f"{result.time_us:.2f}",
+                    round(result.mem_bw_gbps, 3),
+                    fastest_version,
+                    speedup_str,
+                ]
+            )
+
+    print(tabulate(rows, headers=headers))
+
+
+def main():
+    torch.random.manual_seed(123)
+    configs = get_configs()
+    results = []
+    for config in tqdm(configs):
+        result = run_experiment(config)
+        results.append(Experiment(config=config, result=result))
+
+    # Use Tabulate to print results
+    print_results(results)
+
+
+if __name__ == "__main__":
+    main()

dsv3_roofline.png

@@ -709,6 +709,7 @@ def get_extensions():
         mxfp8_sources = [
             os.path.join(mxfp8_extension_dir, "mxfp8_extension.cpp"),
             os.path.join(mxfp8_extension_dir, "mxfp8_cuda.cu"),
+            os.path.join(mxfp8_extension_dir, "mx_block_rearrange_2d_K_groups.cu"),
         ]
 
         # Only add the extension if the source files exist AND we are building for sm100

setup.py

@@ -352,3 +352,67 @@ def test_cuda_mx_dim1_3d_numerics(E, N, K, input_dtype, scaling_mode):
     # Check quantized values
     torch.testing.assert_close(y_d1, y_d1_ref, rtol=0, atol=0)
     assert y_d1.stride() == y_d1_ref.stride(), "quantized tensor strides do not match"
+
+
+@pytest.mark.skipif(
+    not is_sm_at_least_100(),
+    reason="MXFP8 requires CUDA capability 10.0 or greater",
+)
+@pytest.mark.parametrize("m", [256, 512, 1024, 5120])
+@pytest.mark.parametrize("total_k", [512, 1024, 2048, 4096, 8192, 16384])
+@pytest.mark.parametrize("n_groups", [1, 4, 8, 16])
+def test_cuda_mx_block_rearrange_2d_K_groups(
+    m: int,
+    total_k: int,
+    n_groups: int,
+):
+    """
+    Test CUDA kernel for mx_block_rearrange_2d_K_groups against Triton reference.
+    This kernel rearranges E8M0 scales to block-scaled swizzle format for cuBLAS Tmem.
+    """
+    from torchao.prototype.moe_training.kernels.mxfp8.quant import (
+        mx_block_rearrange_2d_K_groups_cuda,
+    )
+
+    device = "cuda"
+    block_size = 32
+    input_data = torch.randn(m, total_k, device=device)
+
+    e8m0_scales, _ = to_mx(
+        input_data, elem_dtype=torch.float8_e4m3fn, block_size=block_size
+    )
+
+    # Generate group end offsets along total_K, then divide by block_size to get scale group end offsets
+    input_group_offsets = generate_jagged_offs(
+        n_groups, total_k, multiple_of=block_size, device=device
+    )
+    scale_group_offsets = input_group_offsets // block_size
+
+    # Triton reference implementation
+    triton_out_scales = triton_mx_block_rearrange_2d_K_groups(
+        e8m0_scales,
+        scale_group_offsets,
+    )
+
+    # CUDA kernel implementation
+    cuda_out_scales = mx_block_rearrange_2d_K_groups_cuda(
+        e8m0_scales,
+        scale_group_offsets,
+    )
+
+    # Check that outputs match
+    assert torch.equal(triton_out_scales, cuda_out_scales.view(torch.float8_e8m0fnu)), (
+        "CUDA and Triton blocked scales not equal"
+    )
+
+    # Verify output shape
+    expected_rows = ((m + 127) // 128) * 128  # Padded to multiple of 128
+    expected_cols = (
+        e8m0_scales.size(1) + n_groups * 4
+    )  # Original cols + padding per group
+    assert cuda_out_scales.shape == (
+        expected_rows,
+        expected_cols,
+    ), (
+        f"Output shape mismatch: expected {(expected_rows, expected_cols)}, got {cuda_out_scales.shape}"
+    )