add differentiable mxfp8 grouped gemm with dynamic quant (forward pass)

stack-info: PR: #2627, branch: danielvegamyhre/stack/23

Author: danielvegamyhre

test/prototype/moe_training/test_scaled_grouped_mm.py

@@ -249,3 +249,25 @@ def test_emulate_mxfp8_grouped_gemm(M, K, N, num_experts):
     sqnr = compute_error(ref_out, out)
     min_sqnr = 27.0
     assert sqnr >= min_sqnr, f"sqnr {sqnr} is too low, must be >= {min_sqnr}"
+
+
+@pytest.mark.parametrize("M", (1024, 4096))
+@pytest.mark.parametrize("K", (1024, 4096))
+@pytest.mark.parametrize("N", (1024, 4096))
+@pytest.mark.parametrize("num_experts", (1, 8, 16))
+def test_mxfp8_grouped_gemm_with_dq_fwd(M, K, N, num_experts):
+    from torchao.prototype.moe_training.scaled_grouped_mm import (
+        _MXFP8GroupedMM,
+    )
+
+    x = torch.randn(M, K, dtype=torch.bfloat16, device="cuda")
+    w_t = torch.randn(num_experts, K, N, 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()
+    block_size = 32
+
+    out = _MXFP8GroupedMM.apply(x, w_t, offs, block_size, torch.bfloat16)
+    ref_out = torch._grouped_mm(x_ref, w_t_ref, offs=offs_ref, out_dtype=torch.bfloat16)
+    sqnr = compute_error(ref_out, out)
+    min_sqnr = 27.0
+    assert sqnr >= min_sqnr, f"sqnr {sqnr} is too low, must be >= {min_sqnr}"

torchao/prototype/moe_training/scaled_grouped_mm.py

@@ -5,7 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import logging
-from typing import Optional
+from typing import Optional, Tuple
 
 import torch
 
@@ -18,6 +18,7 @@
 from torchao.prototype.moe_training.utils import (
     _is_column_major,
 )
+from torchao.prototype.mx_formats.mx_tensor import to_mx
 
 logger: logging.Logger = logging.getLogger(__name__)
 
@@ -268,6 +269,81 @@ def backward(ctx, grad_output: torch.Tensor):
         return grad_A, grad_B.transpose(-2, -1), None, None, None, None
 
 
+class _MXFP8GroupedMM(torch.autograd.Function):
+    """Differentiable implementation of grouped GEMM with dynamic mxpf8 quantization."""
+
+    @staticmethod
+    def forward(
+        ctx,
+        A: torch.Tensor,
+        B_t: torch.Tensor,
+        offs: Optional[torch.Tensor] = None,
+        block_size: int = 32,
+        out_dtype: Optional[torch.dtype] = torch.bfloat16,
+        emulated: bool = True,
+    ) -> torch.Tensor:
+        # torchao _scaled_grouped_mm only supports A=2D and B=3D.
+        assert A.ndim == 2, "A must be 2D"
+        assert B_t.ndim == 3, "B must be 3D"
+        assert block_size == 32, "Only block_size=32 is supported"
+        assert emulated, "Only emulated mxfp8 grouped gemm is supported"
+
+        # Cast to mxpf8 across dim -1.
+        # A_mx shape: (M, K)
+        # A_scale shape: (M, K//block_size)
+        A_scale, A_mx = to_mx(A, elem_dtype=torch.float8_e4m3fn, block_size=block_size)
+
+        # Cast B_t per-expert to mxfp8 across dim1.
+        # B_t_mx shape: (E, K, N)
+        # B_t_scale shape: (E, K//block_size, N)
+        B_t_scale, B_t_mx = _to_mxfp8_3d_expert_weights_dim1(B_t, block_size=block_size)
+
+        # Store what we need for backward.
+        ctx.save_for_backward(A, B_t, offs)
+        ctx.out_dtype = out_dtype
+
+        # Perform scaled grouped GEMM and return result.
+        # output = input @ weight.T
+        # output shape: (M, N)
+        out = emulated_mxfp8_scaled_grouped_mm(
+            A_mx,
+            A_scale,
+            B_t_mx,
+            B_t_scale,
+            offs=offs,
+            block_size=block_size,
+            out_dtype=out_dtype,
+        )
+        return out
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor):
+        raise NotImplementedError
+
+
+def _to_mxfp8_3d_expert_weights_dim1(
+    w_t: torch.Tensor,  # (num_experts, K, N)
+    block_size: int = 32,
+    elem_dtype: torch.dtype = torch.float8_e4m3fn,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Convert a 3D tensor of shape (experts, K, N) to MXFP8 format along dim1.
+    Args:
+        x (torch.Tensor): Input tensor to be converted.
+        block_size (int): Block size for MXFP8 quantization.
+        elem_dtype (torch.dtype): Element dtype for MXFP8 quantization.
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Converted tensor and scale tensor.
+            - scale shape: (expets, K // block_size, N)
+            - output shape: (experts, K, N)
+    """
+    # 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(), elem_dtype=elem_dtype, block_size=block_size
+    )
+    w_t_scale, w_t_mx = w_scale.transpose(-2, -1), w_mx.transpose(-2, -1)
+    return w_t_scale, w_t_mx
+
+
 def emulated_mxfp8_scaled_grouped_mm(
     A_mx: torch.Tensor,
     A_scale: torch.Tensor,