roofline estimator: add float8 rowwise and mxfp8 recipe support (#1789)

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

* Update

[ghstack-poisoned]

Author: vkuzo

benchmarks/float8/float8_roofline.py

@@ -47,6 +47,7 @@
 import pandas as pd
 import sympy
 import torch
+import torch.nn as nn
 import torch.utils.benchmark as benchmark
 import tqdm
 from torch.profiler import ProfilerActivity, profile
@@ -57,8 +58,11 @@
 )
 
 from torchao.float8 import (
+    Float8LinearConfig,
     convert_to_float8_training,
 )
+from torchao.prototype.mx_formats.config import MXLinearConfig
+from torchao.prototype.mx_formats.mx_linear import swap_linear_with_mx_linear
 from torchao.testing.float8.roofline_utils import (
     get_float8_mem_sympy,
     get_gemm_time_sympy,
@@ -93,17 +97,19 @@ def benchmark_fn_in_sec(f, *args, **kwargs):
     return measurement.mean
 
 
-def get_gpu_kernel_time(m, x):
+def get_gpu_kernel_time(m, x, grad_output):
     # warm up
     for _ in range(2):
-        m(x).sum().backward()
+        y = m(x)
+        y.backward(grad_output)
 
     # capture a profiling run
     activities = [ProfilerActivity.CPU, ProfilerActivity.CUDA]
     n_iter = 5
     with profile(activities=activities) as prof:
         for _ in range(n_iter):
-            m(x).sum().backward()
+            y = m(x)
+            y.backward(grad_output)
             torch.cuda.synchronize()
     # get the gpu kernel time and aggregate it
     num_leaf_tensors = 1 + len(list(m.parameters()))
@@ -114,10 +120,28 @@ def get_gpu_kernel_time(m, x):
     return total_time_s
 
 
-def get_gemm_times(M, K, N, fast_accum, cache_filename=None):
+def get_gemm_times(
+    gemm_role: str,
+    M: int,
+    K: int,
+    N: int,
+    fast_accum: bool,
+    bf16_memory_formats: str,
+    float8_recipe_name: Optional[str],
+    mx_recipe_name: Optional[str],
+    cache_filename=None,
+):
+    assert gemm_role in ("output", "grad_input", "grad_weight"), "unsupported"
+    assert bf16_memory_formats in (
+        "row_major:col_major",
+        "row_major:row_major",
+        "col_major:row_major",
+    ), "unsupported"
+
     # Note: this is definitely not the best way to build a cache,
     # but it will do for now.
     if cache_filename is not None:
+        assert False, "TODO retest this for new arguments"
         if os.path.isfile(cache_filename):
             # cache already exists, use it
             with open(cache_filename, "r") as f:
@@ -127,30 +151,48 @@ def get_gemm_times(M, K, N, fast_accum, cache_filename=None):
             cache = dict()
     else:
         cache = dict()
-    key = f"{M},{K},{N},{fast_accum}"
+    key = f"{M},{K},{N},{fast_accum},{bf16_memory_formats}"
     if key in cache:
         return cache[key]
 
     device = torch.device("cuda")
 
     # bf16 time
     x_bf16 = torch.randn(M, K, dtype=torch.bfloat16, device=device)
-    w_bf16 = torch.randn(K, N, dtype=torch.bfloat16, device=device).t().contiguous().t()
+    # w_bf16 = torch.randn(K, N, dtype=torch.bfloat16, device=device).t().contiguous().t()
+    w_bf16 = torch.randn(K, N, dtype=torch.bfloat16, device=device)
+
+    if bf16_memory_formats == "row_major:col_major":
+        w_bf16 = w_bf16.t().contiguous().t()
+    elif bf16_memory_formats == "col_major:row_major":
+        x_bf16 = x_bf16.t().contiguous().t()
+    elif bf16_memory_formats == "col_major:row_major":
+        x_bf16 = x_bf16.t().contiguous().t()
+
     bf16_time_s = get_gpu_kernel_gemm_time_s(torch.mm, x_bf16, w_bf16)
 
     # f8 time
-    d1, d2, d3 = torch.float8_e4m3fn, torch.float8_e4m3fn, torch.bfloat16
-    A = torch.zeros(M, K, device=device, dtype=d1)
-    B = torch.zeros(K, N, device=device, dtype=d2).t().contiguous().t()
-    scale_a = torch.tensor([1.0], device=device)
-    scale_b = torch.tensor([1.0], device=device)
-
-    def do_matmul(A, B):
-        return torch._scaled_mm(
-            A, B, scale_a, scale_b, out_dtype=d3, use_fast_accum=fast_accum
-        )
+    if float8_recipe_name == "rowwise_with_gw_hp" and gemm_role == "grad_weight":
+        f8_time_s = bf16_time_s
+    else:
+        d1, d2, d3 = torch.float8_e4m3fn, torch.float8_e4m3fn, torch.bfloat16
+        A = torch.zeros(M, K, device=device, dtype=d1)
+        B = torch.zeros(K, N, device=device, dtype=d2).t().contiguous().t()
+        if float8_recipe_name == "tensorwise":
+            scale_a = torch.tensor([1.0], device=device)
+            scale_b = torch.tensor([1.0], device=device)
+        elif float8_recipe_name in ("rowwise", "rowwise_with_gw_hp"):
+            scale_a = torch.ones(M, 1, device=device)
+            scale_b = torch.ones(1, N, device=device)
+        else:
+            assert False, "TODO add mx gemm here"
+
+        def do_matmul(A, B):
+            return torch._scaled_mm(
+                A, B, scale_a, scale_b, out_dtype=d3, use_fast_accum=fast_accum
+            )
 
-    f8_time_s = get_gpu_kernel_gemm_time_s(do_matmul, A, B)
+        f8_time_s = get_gpu_kernel_gemm_time_s(do_matmul, A, B)
 
     # save to cache if needed
     if cache_filename is not None:
@@ -164,33 +206,52 @@ def do_matmul(A, B):
 def run(
     outfile: str,
     do_benchmarks: bool = True,
-    shape_gen_name: str = "square",
+    shape_gen_name: str = "pow2",
     gemm_cache_filename: Optional[str] = None,
     n_limit: Optional[int] = None,
+    float8_recipe_name: Optional[str] = None,
+    mx_recipe_name: Optional[str] = None,
+    enable_fusion_modeling: bool = False,
 ):
     """
     Args:
     * `do_benchmarks`: if True, gemm and e2e fwd+bwd of LNLinearSigmoid are benchmarked
-    * `shape_gen_name`: `llama`, `square`, or `sweep`
+    * `shape_gen_name`: `llama`, `pow2`, `pow2_extended`, or `sweep`
     * `gemm_cache_filename (optional)`: file to cache gemm benchmark results
     * `n_limit (optional)`: if specified, only runs `n_limit` iterations
+    * `enable_fusion_modeling`: if False uses Linear, if True uses LNLinearSigmoid and models the fusion of float8 overhead
     """
 
+    assert not (
+        (float8_recipe_name is not None) and (mx_recipe_name is not None)
+    ), "unsupported"
+    if float8_recipe_name is None and mx_recipe_name is None:
+        float8_recipe_name = "tensorwise"
+
+    print(f"GPU: {torch.cuda.get_device_name(0)}")
     print(f"do_benchmarks: {do_benchmarks}")
     print(f"shape_gen_name: {shape_gen_name}")
+    print(f"float8_recipe_name: {float8_recipe_name}")
+    print(f"mx_recipe_name: {mx_recipe_name}")
+    print(f"enable_fusion_modeling: {enable_fusion_modeling}")
 
     M, K, N = sympy.symbols("M K N")
 
-    fp8_mem_time_sympy_dyn_nolimit = get_float8_mem_sympy(
+    fp8_ovhd_time_sympy = get_float8_mem_sympy(
         M,
         K,
         N,
+        float8_recipe_name,
+        mx_recipe_name,
+        enable_fusion_modeling,
+    )
+    bf16_gemm_time_sympy = get_gemm_time_sympy(M, K, N, torch.bfloat16, None, None)
+    fp8_gemm_time_sympy = get_gemm_time_sympy(
+        M, K, N, torch.float8_e4m3fn, float8_recipe_name, mx_recipe_name
     )
-
-    bf16_gemm_time_sympy = get_gemm_time_sympy(M, K, N, torch.bfloat16)
     print("bf16_gemm_time_sympy", bf16_gemm_time_sympy)
-    fp8_gemm_time_sympy = get_gemm_time_sympy(M, K, N, torch.float8_e4m3fn)
     print("fp8_gemm_time_sympy", fp8_gemm_time_sympy)
+    print("fp8_ovhd_time_sympy", fp8_ovhd_time_sympy)
     print()
 
     headers = [
@@ -217,6 +278,9 @@ def run(
         # the difference is the fwd+bwd ln and sigmoid terms, for now to keep things simple
         # we don't break them out and don't have a roofline for them.
         "b_fp8_e2e_spdp",
+        # how well benchmarked gemms match roofline predicted gemms
+        "rb_bf16_gemm_ratio",
+        "rb_fp8_gemm_ratio",
     ]
     results = []
 
@@ -237,43 +301,96 @@ def run(
 
         # if enabled, also measured observed gemm time
         b_bf16_gemm_time_s, b_fp8_gemm_time_s = 0, 0
+        rb_bf16_gemm_ratio = -1
+        rb_fp8_gemm_ratio = -1
+
         if do_benchmarks:
+            # TODO(future): make the bf16 gemm times exactly match the e2e
+            # benchmarks, there is a slight deviation, probably related to gemm
+            # operand memory formats/transpositions below not exactly matching
+            # what PyTorch core is doing for `torch.mm`
+            # input @ weight_t = output
             bf16_g1, f8_g1 = get_gemm_times(
-                M_val, K_val, N_val, True, gemm_cache_filename
+                "output",
+                M_val,
+                K_val,
+                N_val,
+                True,
+                "row_major:col_major",
+                float8_recipe_name,
+                mx_recipe_name,
+                gemm_cache_filename,
             )
+            # grad_output @ weight = grad_input
             bf16_g2, f8_g2 = get_gemm_times(
-                M_val, N_val, K_val, False, gemm_cache_filename
+                "grad_input",
+                M_val,
+                N_val,
+                K_val,
+                False,
+                "row_major:row_major",
+                float8_recipe_name,
+                mx_recipe_name,
+                gemm_cache_filename,
             )
+            # input_t @ grad_output = grad_weight
             bf16_g3, f8_g3 = get_gemm_times(
-                K_val, M_val, N_val, False, gemm_cache_filename
+                "grad_weight",
+                K_val,
+                M_val,
+                N_val,
+                False,
+                "col_major:row_major",
+                float8_recipe_name,
+                mx_recipe_name,
+                gemm_cache_filename,
             )
             b_bf16_gemm_time_s = bf16_g1 + bf16_g2 + bf16_g3
             b_fp8_gemm_time_s = f8_g1 + f8_g2 + f8_g3
+            rb_bf16_gemm_ratio = r_bf16_gemm_time_s / b_bf16_gemm_time_s
+            rb_fp8_gemm_ratio = r_fp8_gemm_time_s / b_fp8_gemm_time_s
 
         # note: cast from sympy.core.numbers.Float to float to make pandas formatting work
         r_fp8_ovhd_time_s = float(
-            fp8_mem_time_sympy_dyn_nolimit.subs(M, M_val).subs(K, K_val).subs(N, N_val)
+            fp8_ovhd_time_sympy.subs(M, M_val).subs(K, K_val).subs(N, N_val)
         )
 
         b_bf16_e2e_time_s, b_fp8_e2e_time_s = 0, 0
         if do_benchmarks:
             # create the model
-            m_orig = LNLinearSigmoid(K_val, N_val).cuda().bfloat16()
+            if enable_fusion_modeling:
+                m_orig = LNLinearSigmoid(K_val, N_val).cuda().bfloat16()
+            else:
+                m_orig = (
+                    nn.Sequential(nn.Linear(K_val, N_val, bias=False)).cuda().bfloat16()
+                )
             x = torch.randn(
                 M_val, K_val, dtype=torch.bfloat16, device="cuda"
             ).requires_grad_()
 
+            # get the gradient of the right shape
+            grad_output = torch.randn(N_val, K_val, dtype=torch.bfloat16, device="cuda")
+
             # get the bf16 gpu kernel time
             torch._dynamo.reset()
             m_bf16 = torch.compile(copy.deepcopy(m_orig))
-            b_bf16_e2e_time_s = get_gpu_kernel_time(m_bf16, x)
+            b_bf16_e2e_time_s = get_gpu_kernel_time(m_bf16, x, grad_output)
 
             # get the float8 dynamic scaling gpu kernel time
 
             torch._dynamo.reset()
-            m_fp8_dyn = convert_to_float8_training(copy.deepcopy(m_orig))
+            if float8_recipe_name is not None:
+                config = Float8LinearConfig.from_recipe_name(float8_recipe_name)
+                m_fp8_dyn = convert_to_float8_training(
+                    copy.deepcopy(m_orig), config=config
+                )
+            else:
+                assert mx_recipe_name is not None
+                config = MXLinearConfig.from_recipe_name(mx_recipe_name)
+                m_fp8_dyn = copy.deepcopy(m_orig)
+                swap_linear_with_mx_linear(m_fp8_dyn, config=config)
             m_fp8_dyn = torch.compile(m_fp8_dyn)
-            b_fp8_e2e_time_s = get_gpu_kernel_time(m_fp8_dyn, x)
+            b_fp8_e2e_time_s = get_gpu_kernel_time(m_fp8_dyn, x, grad_output)
 
         results.append(
             [
@@ -295,6 +412,9 @@ def run(
                 b_bf16_e2e_time_s,
                 b_fp8_e2e_time_s,
                 b_bf16_e2e_time_s / (b_fp8_e2e_time_s + 1e-20),
+                # gemm ratios
+                rb_bf16_gemm_ratio,
+                rb_fp8_gemm_ratio,
             ]
         )
 

benchmarks/float8/utils.py

@@ -152,18 +152,32 @@ def get_name_to_shapes_iter(
         }
         return name_to_shapes_70b.items()
 
-    elif shape_gen_name == "square":
+    elif shape_gen_name == "pow2":
         assert (
             M == K == N == None
         ), f"M, K, N arguments not supported for shape_gen_name {shape_gen_name}"
         name_to_shapes = {}
-        min_power_of_2 = 8  # 256
-        max_power_of_2 = 15  # 32,768
+        min_power_of_2 = 10  # 1024
+        max_power_of_2 = 14  # 16,384
         for idx, power_of_2 in enumerate(range(min_power_of_2, max_power_of_2 + 1)):
             val = 2**power_of_2
             name_to_shapes[idx] = val, val, val
         return name_to_shapes.items()
 
+    elif shape_gen_name == "pow2_extended":
+        assert (
+            M == K == N == None
+        ), f"M, K, N arguments not supported for shape_gen_name {shape_gen_name}"
+        name_to_shapes = {}
+        min_power_of_2 = 10  # 1024
+        max_power_of_2 = 14  # 16,384
+        for idx, power_of_2 in enumerate(range(min_power_of_2, max_power_of_2 + 1)):
+            val1 = 2**power_of_2
+            name_to_shapes[idx * 2] = val1, val1, val1
+            val2 = 2**power_of_2 + 2 ** (power_of_2 - 1)
+            name_to_shapes[idx * 2 + 1] = val2, val2, val2
+        return name_to_shapes.items()
+
     elif shape_gen_name == "sweep":
         assert (
             M == K == N == None