FP8 Grouped Gemm Optimization

Summary:
X-link: facebookresearch/FBGEMM#731

While optimizing MOE, we found that small overheads were a major bottleneck for grouped gemm performance. This diff tackles a few of them, specifically overhead from torch.dynamo wrapping `quantize_fp8_row` and having to slice input tensors before calling `f8f8bf16_rowwise_grouped`.

To fix the former, we enable `triton_quantize_fp8_row` to be directly called, skipping dynamo compatibility. In cases where AOTI isnt needed, this removes a bit of overhead.

To fix the latter, we templatize f8f8fbf16_rowwise_grouped_dynamic to accept at::Tensor instead of lists. We introduce a new wrapper called f8f8bf16_rowwise_grouped_stacked to maintain the behavior where zero_start_index_M isnt provided but a user wants a single contiguous output tensor.

In microbenchmarks, we've found these seemingly small changes can improve TFLOPs by 2X for small workloads.

Reviewed By: jiawenliu64

Differential Revision: D69072529

Author: jwfromm

fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py

@@ -2447,6 +2447,13 @@ def triton_quantize_fp8_row(
         torch.Tensor: fp8 scaled tensor.
         torch.Tensor: reciprocal scale tensor per row.
     """
+    assert a.dim() <= 4, "Triton only supports up to 4 dimension input tensor."
+    a_shape = a.shape
+    while a.dim() < 4:
+        a = a.unsqueeze(0)
+    if zero_start_index_M is not None:
+        # There should be one value of zero_start_index_M per NxK matrix.
+        zero_start_index_M = zero_start_index_M.view(a.shape[0], a.shape[1])
     # Get constant values.
     pt_dtype, tl_dtype, max_fp8, eps = get_fp8_constants()
     num_rows = a.numel() // a.shape[-1]
@@ -2484,7 +2491,7 @@ def triton_quantize_fp8_row(
         USE_INT64=use_int64,
     )
 
-    return a_fp8, a_scale
+    return a_fp8.view(a_shape), a_scale.view(a_shape[:-1])
 
 
 @torch.library.custom_op("triton::quantize_fp8_row", mutates_args=())
@@ -2514,17 +2521,7 @@ def quantize_fp8_row(
         logger.info("Triton does not support cpu, falling back to torch ops.")
         use_triton = False
     if use_triton:
-        assert (
-            a.dim() <= 4
-        ), "Only up to 4 dimension input tensor is supported if use_triton is True"
-        a_shape = a.shape
-        while a.dim() < 4:
-            a = a.unsqueeze(0)
-        if zero_start_index_M is not None:
-            # There should be one value of zero_start_index_M per NxK matrix.
-            zero_start_index_M = zero_start_index_M.view(a.shape[0], a.shape[1])
-        a_fp8, a_scale = triton_quantize_fp8_row(a, scale_ub, zero_start_index_M)
-        return a_fp8.view(a_shape), a_scale.view(a_shape[:-1])
+        return triton_quantize_fp8_row(a, scale_ub, zero_start_index_M)
     # else use pytorch implementation.
     if not output_device:
         output_device = a.device

fbgemm_gpu/experimental/gen_ai/bench/quantize_ops.py

@@ -541,16 +541,11 @@ def compute(self, xq, wq, x_scale, w_scale, m_values, kernel_name=None):
             )
         else:
             # Break tensor into groups, simulates what is done e2e.
-            B = xq.shape[0]
-            xq_group = [xq[i, :, :] for i in range(B)]
-            x_scale_group = [x_scale[i, :] for i in range(B)]
-            wq_group = [wq[i, :, :] for i in range(B)]
-            w_scale_group = [w_scale[i, :] for i in range(B)]
             return torch.ops.fbgemm.f8f8bf16_rowwise_grouped_dynamic(
-                xq_group,
-                wq_group,
-                x_scale_group,
-                w_scale_group,
+                xq,
+                wq,
+                x_scale,
+                w_scale,
                 zero_start_index_M=m_values,
                 kernel_name=kernel_name,
             )