[mxfp8 moe] add support for fbgemm 2d-3d mx8mx8bf16 grouped gemm

stack-info: PR: #2848, branch: danielvegamyhre/stack/55

Author: danielvegamyhre

test/prototype/moe_training/test_scaled_grouped_mm.py

@@ -230,25 +230,26 @@ def compute_reference_forward(
 @pytest.mark.parametrize("num_experts", (1, 8, 16))
 def test_emulate_mxfp8_grouped_gemm_2d_3d(M, K, N, num_experts):
     x = torch.randn(M, K, dtype=torch.bfloat16, device="cuda")
-    w_t = torch.randn(num_experts, K, N, dtype=torch.bfloat16, device="cuda")
+    w = torch.randn(num_experts, N, K, dtype=torch.bfloat16, device="cuda")
     offs = generate_jagged_offs(num_experts, M)
-    x_ref, w_t_ref, offs_ref = x.clone(), w_t.clone(), offs.clone()
+    x_ref, w_ref, offs_ref = x.clone(), w.clone(), offs.clone()
 
     # Quantize inputs to mxpf8 for emulated mxfp8 scaled grouped mm
     block_size = 32
-    x_scale, x_mx = to_mx(x, elem_dtype=torch.float8_e4m3fn, block_size=block_size)
+    x_scale, x_fp8 = to_mx(x, elem_dtype=torch.float8_e4m3fn, block_size=block_size)
 
     # To cast B_t per-expert to mxfp8 across dim1, we transpose the experts, cast along dim -1, then untranspose.
-    w_scale, w_mx = to_mx(
-        w_t.transpose(-2, -1).contiguous(),
+    w_scale, w_fp8 = to_mx(
+        w,
         elem_dtype=torch.float8_e4m3fn,
         block_size=block_size,
     )
-    w_t_scale, w_t_mx = w_scale.transpose(-2, -1), w_mx.transpose(-2, -1)
 
-    ref_out = torch._grouped_mm(x_ref, w_t_ref, offs=offs_ref, out_dtype=torch.bfloat16)
+    ref_out = torch._grouped_mm(
+        x_ref, w_ref.transpose(-2, -1), offs=offs_ref, out_dtype=torch.bfloat16
+    )
     out = _emulated_mxfp8_scaled_grouped_mm_2d_3d(
-        x_mx, x_scale, w_t_mx, w_t_scale, offs=offs, out_dtype=torch.bfloat16
+        x_fp8, x_scale, w_fp8, w_scale, offs=offs, out_dtype=torch.bfloat16
     )
 
     sqnr = compute_error(ref_out, out)
@@ -305,19 +306,27 @@ def test_emulate_mxfp8_grouped_gemm_2d_2d(M, N, num_experts):
 
 
 @skip_if_rocm("ROCm not supported")
-@pytest.mark.parametrize("M,K,N", [(1024, 1024, 1024), (1024, 2048, 4096)])
-@pytest.mark.parametrize("num_experts", (1, 8, 16))
+@pytest.mark.parametrize("M,K,N", [(256, 512, 512)])
+@pytest.mark.parametrize("num_experts", (2,))
 def test_mxfp8_grouped_gemm_with_dq_fwd_bwd(M, K, N, num_experts):
     from torchao.prototype.moe_training.scaled_grouped_mm import (
         _MXFP8GroupedMM,
     )
 
     block_size = 32
     x = torch.randn(M, K, dtype=torch.bfloat16, device="cuda", requires_grad=True)
-    w_t = torch.randn(
-        num_experts, K, N, dtype=torch.bfloat16, device="cuda", requires_grad=True
+    w = torch.randn(
+        num_experts,
+        N,
+        K,
+        dtype=torch.bfloat16,
+        device="cuda",
     )
-    offs = generate_jagged_offs(num_experts, M, multiple_of=block_size)
+    w_t = w.transpose(-2, -1).requires_grad_(True)
+    # TODO: use non-uniform group sizes once kernel supports it
+    # offs = generate_jagged_offs(num_experts, M, multiple_of=block_size)
+    group_size = M // num_experts
+    offs = torch.arange(group_size, M + 1, group_size, device="cuda", dtype=torch.int32)
     x_ref, w_t_ref, offs_ref = (
         x.clone().detach().requires_grad_(True),
         w_t.clone().detach().requires_grad_(True),

test/prototype/moe_training/test_training.py

@@ -129,6 +129,9 @@ def moe_module_filter_fn(mod: nn.Module, cur_fqn: str) -> bool:
         )
 
 
+@pytest.mark.skip(
+    "temporarily disable until non-uniform group sizes are supported by mxfp8 grouped gemm"
+)
 @pytest.mark.parametrize(
     "target_fqns",
     [

torchao/prototype/moe_training/kernels/__init__.py

@@ -7,3 +7,6 @@
 from torchao.prototype.moe_training.kernels.jagged_float8_scales import (
     triton_fp8_per_group_rowwise_scales as triton_fp8_per_group_rowwise_scales,
 )
+from torchao.prototype.moe_training.kernels.mxfp8 import (
+    fbgemm_mxfp8_grouped_mm_2d_3d as fbgemm_mxfp8_grouped_mm_2d_3d,
+)

torchao/prototype/moe_training/kernels/mxfp8.py

@@ -0,0 +1,152 @@
+import logging
+
+import torch
+
+from torchao.prototype.mx_formats.utils import (
+    to_blocked_per_group_2d,
+    to_blocked_per_group_3d,
+)
+
+logger: logging.Logger = logging.getLogger(__name__)
+
+try:
+    import fbgemm_gpu.experimental.gen_ai  # noqa: F401
+except Exception as e:
+    logging.warning(
+        f"fbgemm_gpu_genai package is required for this feature but import failed with exception: {e}"
+        "Please install nightly builds of pytorch and fbgemm_gpu_genai build using this command and try again: "
+        "pip3 install --force-reinstall --pre torch fbgemm-gpu-genai --index-url https://download.pytorch.org/whl/nightly/cu129"
+        "If errors persist, please file a bug report."
+    )
+
+
+@torch.library.custom_op("torchao::fbgemm_mxfp8_grouped_mm_2d_3d", mutates_args={})
+def fbgemm_mxfp8_grouped_mm_2d_3d(
+    A_fp8: torch.Tensor,
+    A_scales: torch.Tensor,
+    B_fp8: torch.Tensor,
+    B_scales: torch.Tensor,
+    offs: torch.Tensor,
+    block_size: int = 32,
+    out_dtype: torch.dtype = torch.bfloat16,
+) -> torch.Tensor:
+    assert A_fp8.ndim == 2, "A_fp8 tensor must be 2D"
+    assert B_fp8.ndim == 3, "B_fp8 tensor must be 3D"
+    assert block_size == 32, "Only block_size=32 is supported"
+    assert out_dtype == torch.bfloat16, "Only out_dtype=bfloat16 is supported"
+    assert A_fp8.shape[-1] == B_fp8.shape[-1], "A_fp8 and B_fp8 must have same last dim"
+
+    # Convert scales for each group to blocked format.
+    Mg, K = A_fp8.shape
+    A_scales_blocked, starting_row_after_padding = to_blocked_per_group_2d(
+        A_scales, offs, Mg, K
+    )
+    B_scales_blocked = to_blocked_per_group_3d(B_scales)
+
+    # From this, we compute `group_sizes` and `starting_row_after_padding`:
+    # group_sizes = [32, 32, 64]
+    # starting_row_after_padding = [0, 32, 64, 128]
+    group_sizes = torch.diff(starting_row_after_padding).to(torch.int64)
+
+    # TODO: remove debug logging once prototype is more mature.
+
+    out = torch.ops.fbgemm.mx8mx8bf16_grouped_stacked(
+        A_fp8,
+        B_fp8,
+        A_scales_blocked,
+        B_scales_blocked,
+        group_sizes,
+        starting_row_after_padding=starting_row_after_padding,
+    )
+    return out
+
+
+@fbgemm_mxfp8_grouped_mm_2d_3d.register_fake
+def _fbgemm_mxfp8_grouped_mm_2d_3d_fake(
+    A_fp8: torch.Tensor,
+    B_fp8: torch.Tensor,
+    A_scales: torch.Tensor,
+    B_scales: torch.Tensor,
+    offs: torch.Tensor,
+) -> torch.Tensor:
+    assert A_fp8.ndim == 2, "A_fp8 tensor must be 2D"
+    assert B_fp8.ndim == 3, "B_fp8 tensor must be 3D"
+    mg, k = A_fp8.shape
+    e, k, n = B_fp8.shape
+    n_groups = offs.numel()
+    assert n_groups == e, (
+        "Size of `offs` (number of groups) must match first dim of `B_fp8`"
+    )
+    output = torch.empty((mg, n), dtype=torch.bfloat16, device=A_fp8.device)
+    return output
+
+
+def _log_inputs(
+    A_fp8: torch.Tensor,
+    B_fp8: torch.Tensor,
+    A_scales: torch.Tensor,
+    A_scales_blocked: torch.Tensor,
+    B_scales: torch.Tensor,
+    B_scales_blocked: torch.Tensor,
+    offs: torch.Tensor,
+    group_sizes: torch.Tensor,
+    starting_row_after_padding: torch.Tensor,
+):
+    print(f"offs: {offs}")
+    print("A_fp8.shape", A_fp8.shape, "stride", A_fp8.stride(), "dtype", A_fp8.dtype)
+    print(
+        "B_fp8.shape",
+        B_fp8.shape,
+        "stride",
+        B_fp8.stride(),
+        "dtype",
+        B_fp8.dtype,
+    )
+    print(
+        "A_scales (non-blocked)",
+        A_scales.shape,
+        "stride",
+        A_scales.stride(),
+        "dtype",
+        A_scales.dtype,
+    )
+    print(
+        "A_scaless_blocked.shape",
+        A_scales_blocked.shape,
+        "stride",
+        A_scales_blocked.stride(),
+        "dtype",
+        A_scales_blocked.dtype,
+    )
+    print(
+        "B_scales (non-blocked)",
+        B_scales.shape,
+        "stride",
+        B_scales.stride(),
+        "dtype",
+        B_scales.dtype,
+    )
+    print(
+        "B_scales_blocked.shape",
+        B_scales_blocked.shape,
+        "stride",
+        B_scales_blocked.stride(),
+        "dtype",
+        B_scales_blocked.dtype,
+    )
+    print(
+        "group_sizes",
+        group_sizes,
+        "group_sizes.stride",
+        group_sizes.stride(),
+        "dtype",
+        group_sizes.dtype,
+    )
+    print(
+        "starting_row_after_padding",
+        starting_row_after_padding,
+        "stride",
+        starting_row_after_padding.stride(),
+        "dtype",
+        starting_row_after_padding.dtype,
+    )