sgl-fa4 4.0.1__py3-none-any.whl

sgl_fa4/VERSION

@@ -0,0 +1 @@
+4.0.1

sgl_fa4/__init__.py

@@ -0,0 +1,4 @@
+from pathlib import Path
+
+# you can use os.path and open() as well
+__version__ = Path(__file__).parent.joinpath("VERSION").read_text()

sgl_fa4/cute/__init__.py

@@ -0,0 +1,20 @@
+"""Flash Attention CUTE (CUDA Template Engine) implementation."""
+
+
+import cutlass.cute as cute
+
+from .interface import (
+    flash_attn_func,
+    flash_attn_varlen_func,
+)
+
+from sgl_fa4.cute.cute_dsl_utils import cute_compile_patched
+
+# Patch cute.compile to optionally dump SASS
+cute.compile = cute_compile_patched
+
+
+__all__ = [
+    "flash_attn_func",
+    "flash_attn_varlen_func",
+]

sgl_fa4/cute/ampere_helpers.py

@@ -0,0 +1,103 @@
+# Copyright (c) 2025, Tri Dao.
+from typing import Type, Callable, Optional
+
+import cutlass
+import cutlass.cute as cute
+
+
+def get_smem_layout_atom(dtype: Type[cutlass.Numeric], k_dim: int) -> cute.ComposedLayout:
+    dtype_byte = cutlass.const_expr(dtype.width // 8)
+    bytes_per_row = cutlass.const_expr(k_dim * dtype_byte)
+    smem_k_block_size = (
+        cutlass.const_expr(
+            128
+            if bytes_per_row % 128 == 0
+            else (64 if bytes_per_row % 64 == 0 else (32 if bytes_per_row % 32 == 0 else 16))
+        )
+        // dtype_byte
+    )
+    swizzle_bits = (
+        4
+        if smem_k_block_size == 128
+        else (3 if smem_k_block_size == 64 else (2 if smem_k_block_size == 32 else 1))
+    )
+    swizzle_base = 2 if dtype_byte == 4 else (3 if dtype_byte == 2 else 4)
+    return cute.make_composed_layout(
+        cute.make_swizzle(swizzle_bits, swizzle_base, swizzle_base),
+        0,
+        cute.make_ordered_layout(
+            (8 if cutlass.const_expr(k_dim % 32 == 0) else 16, smem_k_block_size), order=(1, 0)
+        ),
+    )
+
+
+@cute.jit
+def gemm(
+    tiled_mma: cute.TiledMma,
+    acc: cute.Tensor,
+    tCrA: cute.Tensor,
+    tCrB: cute.Tensor,
+    tCsA: cute.Tensor,
+    tCsB: cute.Tensor,
+    smem_thr_copy_A: cute.TiledCopy,
+    smem_thr_copy_B: cute.TiledCopy,
+    hook_fn: Optional[Callable] = None,
+    A_in_regs: cutlass.Constexpr[bool] = False,
+    B_in_regs: cutlass.Constexpr[bool] = False,
+    swap_AB: cutlass.Constexpr[bool] = False,
+) -> None:
+    if cutlass.const_expr(swap_AB):
+        gemm(
+            tiled_mma,
+            acc,
+            tCrB,
+            tCrA,
+            tCsB,
+            tCsA,
+            smem_thr_copy_B,
+            smem_thr_copy_A,
+            hook_fn,
+            A_in_regs=B_in_regs,
+            B_in_regs=A_in_regs,
+            swap_AB=False,
+        )
+    else:
+        tCrA_copy_view = smem_thr_copy_A.retile(tCrA)
+        tCrB_copy_view = smem_thr_copy_B.retile(tCrB)
+        if cutlass.const_expr(not A_in_regs):
+            cute.copy(smem_thr_copy_A, tCsA[None, None, 0], tCrA_copy_view[None, None, 0])
+        if cutlass.const_expr(not B_in_regs):
+            cute.copy(smem_thr_copy_B, tCsB[None, None, 0], tCrB_copy_view[None, None, 0])
+        for k in cutlass.range_constexpr(cute.size(tCsA.shape[2])):
+            if k < cute.size(tCsA.shape[2]) - 1:
+                if cutlass.const_expr(not A_in_regs):
+                    cute.copy(
+                        smem_thr_copy_A, tCsA[None, None, k + 1], tCrA_copy_view[None, None, k + 1]
+                    )
+                if cutlass.const_expr(not B_in_regs):
+                    cute.copy(
+                        smem_thr_copy_B, tCsB[None, None, k + 1], tCrB_copy_view[None, None, k + 1]
+                    )
+            cute.gemm(tiled_mma, acc, tCrA[None, None, k], tCrB[None, None, k], acc)
+            if cutlass.const_expr(k == 0 and hook_fn is not None):
+                hook_fn()
+
+
+@cute.jit
+def gemm_rs(
+    tiled_mma: cute.TiledMma,
+    acc: cute.Tensor,
+    tCrA: cute.Tensor,
+    tCrB: cute.Tensor,
+    tCsB: cute.Tensor,
+    smem_thr_copy_B: cute.TiledCopy,
+    hook_fn: Optional[Callable] = None,
+) -> None:
+    tCrB_copy_view = smem_thr_copy_B.retile(tCrB)
+    cute.copy(smem_thr_copy_B, tCsB[None, None, 0], tCrB_copy_view[None, None, 0])
+    for k in cutlass.range_constexpr(cute.size(tCrA.shape[2])):
+        if cutlass.const_expr(k < cute.size(tCrA.shape[2]) - 1):
+            cute.copy(smem_thr_copy_B, tCsB[None, None, k + 1], tCrB_copy_view[None, None, k + 1])
+        cute.gemm(tiled_mma, acc, tCrA[None, None, k], tCrB[None, None, k], acc)
+        if cutlass.const_expr(k == 0 and hook_fn is not None):
+            hook_fn()

sgl_fa4/cute/barrier.py

@@ -0,0 +1,71 @@
+import cutlass
+import cutlass.cute as cute
+from cutlass import Int32
+from cutlass.cutlass_dsl import T, dsl_user_op
+from cutlass._mlir.dialects import llvm
+
+
+@dsl_user_op
+def ld_acquire(lock_ptr: cute.Pointer, *, loc=None, ip=None) -> cutlass.Int32:
+    lock_ptr_i64 = lock_ptr.toint(loc=loc, ip=ip).ir_value()
+    state = llvm.inline_asm(
+        T.i32(),
+        [lock_ptr_i64],
+        "ld.global.acquire.gpu.b32 $0, [$1];",
+        "=r,l",
+        has_side_effects=True,
+        is_align_stack=False,
+        asm_dialect=llvm.AsmDialect.AD_ATT,
+    )
+    return cutlass.Int32(state)
+
+
+@dsl_user_op
+def red_relaxed(
+    lock_ptr: cute.Pointer, val: cutlass.Constexpr[Int32], *, loc=None, ip=None
+) -> None:
+    lock_ptr_i64 = lock_ptr.toint(loc=loc, ip=ip).ir_value()
+    llvm.inline_asm(
+        None,
+        [lock_ptr_i64, Int32(val).ir_value(loc=loc, ip=ip)],
+        "red.relaxed.gpu.global.add.s32 [$0], $1;",
+        "l,r",
+        has_side_effects=True,
+        is_align_stack=False,
+        asm_dialect=llvm.AsmDialect.AD_ATT,
+    )
+
+
+@dsl_user_op
+def red_release(
+    lock_ptr: cute.Pointer, val: cutlass.Constexpr[Int32], *, loc=None, ip=None
+) -> None:
+    lock_ptr_i64 = lock_ptr.toint(loc=loc, ip=ip).ir_value()
+    llvm.inline_asm(
+        None,
+        [lock_ptr_i64, Int32(val).ir_value(loc=loc, ip=ip)],
+        "red.release.gpu.global.add.s32 [$0], $1;",
+        "l,r",
+        has_side_effects=True,
+        is_align_stack=False,
+        asm_dialect=llvm.AsmDialect.AD_ATT,
+    )
+
+
+@cute.jit
+def wait_eq(lock_ptr: cute.Pointer, thread_idx: int | Int32, flag_offset: int, val: Int32) -> None:
+    flag_ptr = lock_ptr + flag_offset
+    if thread_idx == 0:
+        read_val = Int32(0)
+        while read_val != val:
+            read_val = ld_acquire(flag_ptr)
+
+
+@cute.jit
+def arrive_inc(
+    lock_ptr: cute.Pointer, thread_idx: int | Int32, flag_offset: int, val: cutlass.Constexpr[Int32]
+) -> None:
+    flag_ptr = lock_ptr + flag_offset
+    if thread_idx == 0:
+        red_release(flag_ptr, val)
+        # red_relaxed(flag_ptr, val)

sgl_fa4/cute/benchmark.py

@@ -0,0 +1,268 @@
+# Copyright (c) 2023, Tri Dao.
+"""Useful functions for writing test code."""
+
+import torch
+import torch.utils.benchmark as benchmark
+
+
+def benchmark_forward(
+    fn, *inputs, repeats=10, desc="", verbose=True, amp=False, amp_dtype=torch.float16, **kwinputs
+):
+    """Use Pytorch Benchmark on the forward pass of an arbitrary function."""
+    if verbose:
+        print(desc, "- Forward pass")
+
+    def amp_wrapper(*inputs, **kwinputs):
+        with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+            fn(*inputs, **kwinputs)
+
+    t = benchmark.Timer(
+        stmt="fn_amp(*inputs, **kwinputs)",
+        globals={"fn_amp": amp_wrapper, "inputs": inputs, "kwinputs": kwinputs},
+        num_threads=torch.get_num_threads(),
+    )
+    m = t.timeit(repeats)
+    if verbose:
+        print(m)
+    return t, m
+
+
+def benchmark_backward(
+    fn,
+    *inputs,
+    grad=None,
+    repeats=10,
+    desc="",
+    verbose=True,
+    amp=False,
+    amp_dtype=torch.float16,
+    **kwinputs,
+):
+    """Use Pytorch Benchmark on the backward pass of an arbitrary function."""
+    if verbose:
+        print(desc, "- Backward pass")
+    with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+        y = fn(*inputs, **kwinputs)
+        if type(y) is tuple:
+            y = y[0]
+    if grad is None:
+        grad = torch.randn_like(y)
+    else:
+        if grad.shape != y.shape:
+            raise RuntimeError("Grad shape does not match output shape")
+
+    def f(*inputs, y, grad):
+        # Set .grad to None to avoid extra operation of gradient accumulation
+        for x in inputs:
+            if isinstance(x, torch.Tensor):
+                x.grad = None
+        y.backward(grad, retain_graph=True)
+
+    t = benchmark.Timer(
+        stmt="f(*inputs, y=y, grad=grad)",
+        globals={"f": f, "inputs": inputs, "y": y, "grad": grad},
+        num_threads=torch.get_num_threads(),
+    )
+    m = t.timeit(repeats)
+    if verbose:
+        print(m)
+    return t, m
+
+
+def benchmark_combined(
+    fn,
+    *inputs,
+    grad=None,
+    repeats=10,
+    desc="",
+    verbose=True,
+    amp=False,
+    amp_dtype=torch.float16,
+    **kwinputs,
+):
+    """Use Pytorch Benchmark on the forward+backward pass of an arbitrary function."""
+    if verbose:
+        print(desc, "- Forward + Backward pass")
+    with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+        y = fn(*inputs, **kwinputs)
+        if type(y) is tuple:
+            y = y[0]
+    if grad is None:
+        grad = torch.randn_like(y)
+    else:
+        if grad.shape != y.shape:
+            raise RuntimeError("Grad shape does not match output shape")
+
+    def f(grad, *inputs, **kwinputs):
+        for x in inputs:
+            if isinstance(x, torch.Tensor):
+                x.grad = None
+        with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+            y = fn(*inputs, **kwinputs)
+            if type(y) is tuple:
+                y = y[0]
+        y.backward(grad, retain_graph=True)
+
+    t = benchmark.Timer(
+        stmt="f(grad, *inputs, **kwinputs)",
+        globals={"f": f, "fn": fn, "inputs": inputs, "grad": grad, "kwinputs": kwinputs},
+        num_threads=torch.get_num_threads(),
+    )
+    m = t.timeit(repeats)
+    if verbose:
+        print(m)
+    return t, m
+
+
+def benchmark_fwd_bwd(
+    fn,
+    *inputs,
+    grad=None,
+    repeats=10,
+    desc="",
+    verbose=True,
+    amp=False,
+    amp_dtype=torch.float16,
+    **kwinputs,
+):
+    """Use Pytorch Benchmark on the forward+backward pass of an arbitrary function."""
+    return (
+        benchmark_forward(
+            fn,
+            *inputs,
+            repeats=repeats,
+            desc=desc,
+            verbose=verbose,
+            amp=amp,
+            amp_dtype=amp_dtype,
+            **kwinputs,
+        ),
+        benchmark_backward(
+            fn,
+            *inputs,
+            grad=grad,
+            repeats=repeats,
+            desc=desc,
+            verbose=verbose,
+            amp=amp,
+            amp_dtype=amp_dtype,
+            **kwinputs,
+        ),
+    )
+
+
+def benchmark_all(
+    fn,
+    *inputs,
+    grad=None,
+    repeats=10,
+    desc="",
+    verbose=True,
+    amp=False,
+    amp_dtype=torch.float16,
+    **kwinputs,
+):
+    """Use Pytorch Benchmark on the forward+backward pass of an arbitrary function."""
+    return (
+        benchmark_forward(
+            fn,
+            *inputs,
+            repeats=repeats,
+            desc=desc,
+            verbose=verbose,
+            amp=amp,
+            amp_dtype=amp_dtype,
+            **kwinputs,
+        ),
+        benchmark_backward(
+            fn,
+            *inputs,
+            grad=grad,
+            repeats=repeats,
+            desc=desc,
+            verbose=verbose,
+            amp=amp,
+            amp_dtype=amp_dtype,
+            **kwinputs,
+        ),
+        benchmark_combined(
+            fn,
+            *inputs,
+            grad=grad,
+            repeats=repeats,
+            desc=desc,
+            verbose=verbose,
+            amp=amp,
+            amp_dtype=amp_dtype,
+            **kwinputs,
+        ),
+    )
+
+
+def pytorch_profiler(
+    fn,
+    *inputs,
+    trace_filename=None,
+    backward=False,
+    amp=False,
+    amp_dtype=torch.float16,
+    cpu=False,
+    verbose=True,
+    **kwinputs,
+):
+    """Wrap benchmark functions in Pytorch profiler to see CUDA information."""
+    if backward:
+        with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+            out = fn(*inputs, **kwinputs)
+            if type(out) is tuple:
+                out = out[0]
+            g = torch.randn_like(out)
+    for _ in range(30):  # Warm up
+        if backward:
+            for x in inputs:
+                if isinstance(x, torch.Tensor):
+                    x.grad = None
+        with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+            out = fn(*inputs, **kwinputs)
+            if type(out) is tuple:
+                out = out[0]
+        # Backward should be done outside autocast
+        if backward:
+            out.backward(g, retain_graph=True)
+    activities = ([torch.profiler.ProfilerActivity.CPU] if cpu else []) + [
+        torch.profiler.ProfilerActivity.CUDA
+    ]
+    with torch.profiler.profile(
+        activities=activities,
+        record_shapes=True,
+        # profile_memory=True,
+        with_stack=True,
+    ) as prof:
+        if backward:
+            for x in inputs:
+                if isinstance(x, torch.Tensor):
+                    x.grad = None
+        with torch.autocast(device_type="cuda", dtype=amp_dtype, enabled=amp):
+            out = fn(*inputs, **kwinputs)
+            if type(out) is tuple:
+                out = out[0]
+        if backward:
+            out.backward(g, retain_graph=True)
+    if verbose:
+        # print(prof.key_averages().table(sort_by="self_cuda_time_total", row_limit=50))
+        print(prof.key_averages().table(row_limit=50))
+    if trace_filename is not None:
+        prof.export_chrome_trace(trace_filename)
+
+
+def benchmark_memory(fn, *inputs, desc="", verbose=True, **kwinputs):
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.synchronize()
+    fn(*inputs, **kwinputs)
+    torch.cuda.synchronize()
+    mem = torch.cuda.max_memory_allocated() / ((2**20) * 1000)
+    if verbose:
+        print(f"{desc} max memory: {mem}GB")
+    torch.cuda.empty_cache()
+    return mem