IQ.Pilot Release Commit @ 0798119

tinygrad_repo/extra/llama_kernels/__init__.py

@@ -0,0 +1,52 @@
+from __future__ import annotations
+import functools, pathlib
+from dataclasses import replace
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import shape_to_shape_arg
+from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
+
+FP8_MAX = 448.0
+NUM_WG, THREADS_PER_WG = 1024, 256
+
+# per-device abs max without allreduce
+@functools.cache
+def _local_abs_max_fxn(x_p, device):
+  x = Tensor(x_p, device=device)
+  inner = Tensor(x.uop.replace(src=(shape_to_shape_arg(x.uop.shard_shape),), arg=replace(x.uop.arg, axis=None))) if x.uop.axis is not None else x
+  return (inner.abs().max(),)
+
+def local_abs_max(x:Tensor) -> Tensor:
+  param = x.as_param(0)
+  fxn = _local_abs_max_fxn(param.uop, x.device)
+  return Tensor(fxn[0].uop.call(x.uop).gettuple(0))
+
+def scalar_amax(amax_buf:Tensor) -> Tensor:
+  if isinstance(amax_buf.device, tuple):
+    return local_abs_max(amax_buf).detach()
+  return amax_buf.max().detach()
+
+def shard_shape(shape:tuple, axis:int, ndev:int) -> list:
+  s = list(shape)
+  s[axis] //= ndev
+  return s
+
+def dname_of(device) -> str:
+  if isinstance(device, tuple): return device[0].split(":")[0]
+  return device.split(":")[0] if isinstance(device, str) else device
+
+def alloc_like(shape, dtype, device, axis=None) -> Tensor:
+  if isinstance(device, tuple) and axis is not None:
+    return Tensor(Tensor.invalids(*shard_shape(shape, axis, len(device)), dtype=dtype, device=device).uop.multi(axis), device=device)
+  return Tensor.invalids(*shape, dtype=dtype, device=device)
+
+def alloc_local(shape, dtype, device, axis=None) -> Tensor:
+  if isinstance(device, tuple) and axis is not None:
+    return Tensor(Tensor.invalids(*shape, dtype=dtype, device=device).uop.multi(0), device=device)
+  return Tensor.invalids(*shape, dtype=dtype, device=device)
+
+def compile_hip(src:str, defines:list[str]):
+  return HIPCCCompiler("gfx950", ["-std=c++20", "-ffast-math", *defines]).compile_cached(src)
+
+def compile_cpp(cpp_dir:pathlib.Path, cpp_name:str, n_elems:int, hidden:int):
+  src = (cpp_dir/cpp_name).read_text()
+  return src, compile_hip(src, [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={hidden}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"])

tinygrad_repo/extra/llama_kernels/cast_amax/__init__.py

@@ -0,0 +1,76 @@
+from __future__ import annotations
+import functools, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, compile_cpp, alloc_like, alloc_local, scalar_amax, dname_of
+
+# module-level mailbox: grad_xw13 UOp -> (grad_xw13_fp8 UOp, inv_scale UOp)
+# lets cdna_asm_gemm's bwd reuse the fp8 companion produced by the fused silu_mul bwd kernel
+# instead of doing a redundant bf16 -> fp8 quantize.
+_grad_fp8_mailbox:dict[UOp, tuple[UOp, UOp]] = {}
+
+@functools.cache
+def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_buf:UOp,
+                          xw13:UOp, grad_x2:UOp, amax_state:UOp, grad_amax_state:UOp, dname:str) -> UOp:
+  hidden = xw13.shape[2] // 2
+  n_elems = xw13.shape[0] * xw13.shape[1] * hidden
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 * 3 + n_elems * 2 + NUM_WG * 4 + 4
+  sink = UOp.sink(grad_xw13_fp8.base, grad_amax_buf.base,
+                  xw13.base, grad_x2.base, amax_state.base, grad_amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"fused_silu_mul_bwd_w13_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=mem)))
+  src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_bwd_w13.cpp", n_elems, hidden)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
+
+@functools.cache
+def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_buf:UOp, xw13:UOp, amax_state:UOp, grad_amax_state:UOp, dname:str) -> UOp:
+  # NOTE: grad_amax_state is plumbed through as an unused fwd input so the bwd kernel can read it via kernel.src
+  hidden = xw13.shape[2] // 2
+  n_elems = xw13.shape[0] * xw13.shape[1] * hidden
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 * 2 + n_elems + NUM_WG * 4
+  sink = UOp.sink(fp8_out.base, amax_buf.base, xw13.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"fused_silu_mul_cast_amax_w13_{n_elems}", estimates=Estimates(ops=5*n_elems, mem=mem)))
+  src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_fwd_w13.cpp", n_elems, hidden)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
+
+def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):
+  _, _, xw13, amax_state, grad_amax_state = kernel.src[1:]
+  device = xw13.device
+  axis = xw13.axis if isinstance(device, tuple) else None
+  grad_xw13_fp8 = alloc_like(xw13.shape, dtypes.fp8e4m3,  device, axis)
+  grad_amax_buf = alloc_local((NUM_WG,), dtypes.float32,  device, axis)
+  grad_amax_state_t = Tensor(grad_amax_state, device=device)
+  fxn = functools.partial(_custom_fused_bwd_w13, dname=dname_of(device))
+  grad_xw13_fp8, grad_amax_buf, *_ = Tensor.custom_kernel(
+    grad_xw13_fp8, grad_amax_buf,
+    Tensor(xw13, device=device), Tensor(gradient, device=device).cast(dtypes.bfloat16),
+    Tensor(amax_state, device=device), grad_amax_state_t, fxn=fxn)
+  grad_xw13_uop = grad_xw13_fp8.uop.cast(dtypes.bfloat16)
+  inv_scale = (grad_amax_state_t.float() + 1e-8) / FP8_MAX
+  new_grad_amax = scalar_amax(grad_amax_buf)
+  store_effect = grad_amax_state_t.uop.store(new_grad_amax.uop)
+  assert grad_xw13_fp8.uop.op is Ops.AFTER, f"expected AFTER, got {grad_xw13_fp8.uop.op}"
+  grad_xw13_fp8_uop = grad_xw13_fp8.uop.replace(src=grad_xw13_fp8.uop.src + (store_effect,))
+  # Stash fp8 companion for cdna_asm_gemm's bwd to attach to grad_a.
+  _grad_fp8_mailbox[grad_xw13_uop] = (grad_xw13_fp8_uop, inv_scale.uop)
+  return (None, None, grad_xw13_uop, None, None)
+
+def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_state:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns (fp8, inv_scale, new_amax)
+  # grad_amax_state: delayed amax for grad_xw13 fp8 quantization in the backward.
+  assert xw13.dtype == dtypes.bfloat16, f"expected bf16, got {xw13.dtype}"
+  MBS, SEQ, H2 = xw13.shape
+  assert H2 % 2 == 0, f"w13 last-axis must be even, got {H2}"
+  HIDDEN = H2 // 2
+  axis = xw13.uop.axis if isinstance(xw13.device, tuple) else None
+  fp8_out  = alloc_like((MBS, SEQ, HIDDEN), fp8_dtype,      xw13.device, axis)
+  amax_buf = alloc_local((NUM_WG,),         dtypes.float32, xw13.device, axis)
+  fxn = functools.partial(_custom_fused_cast_amax_w13, dname=dname_of(xw13.device))
+  fp8_out, amax_buf, *_ = Tensor.custom_kernel(fp8_out, amax_buf, xw13, amax_state, grad_amax_state,
+                                                fxn=fxn, grad_fxn=_fused_quantize_bwd_w13)
+  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
+  return fp8_out, inv_scale, scalar_amax(amax_buf)

tinygrad_repo/extra/llama_kernels/cast_amax/cast_amax_bwd_w13.cpp

@@ -0,0 +1,91 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+#include <hip/hip_fp8.h>
+
+#ifndef N_ELEMS
+#define N_ELEMS 234881024
+#endif
+#ifndef HIDDEN
+#define HIDDEN 14336
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % VEC == 0, "N_ELEMS must be divisible by VEC");
+static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC");
+
+// fused silu*mul backward, two outputs in a single HBM pass:
+//   1) fp8  grad_xw13_fp8  — delayed-scale quantize using grad_amax_state (mailbox to matmul bwd)
+//   2) fp32 grad_amax_buf  — per-WG partial |grad_xw13|, reduced into next step's grad_amax_state
+// grad_amax_state is read for the fp8 scale. The store of new_grad_amax into grad_amax_state's
+// buffer is built in Python as a separate effect and threaded into grad_a via .after(store).
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_silu_mul_bwd_w13(
+    __hip_fp8_storage_t*  __restrict__ grad_xw13_fp8_out,    // fp8,  2*N_ELEMS
+    float*                __restrict__ grad_amax_buf,        // fp32, NUM_WG per-WG partials
+    const __hip_bfloat16* __restrict__ xw13,                 // bf16, 2*N_ELEMS
+    const __hip_bfloat16* __restrict__ grad_x2,              // bf16, N_ELEMS
+    const float*          __restrict__ amax_state,           // fp32 scalar (fwd x2 amax)
+    const float*          __restrict__ grad_amax_state)      // fp32 scalar (delayed grad amax)
+{
+  __shared__ float sdata[THREADS_PER_WG];
+
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+  const int gid = wg * THREADS_PER_WG + tid;
+  const int stride_elems = NUM_WG * THREADS_PER_WG * VEC;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  const float g_scale = FP8_MAX / (static_cast<float>(*grad_amax_state) + 1e-8f);
+  float local_max = 0.0f;
+
+  for (int base = gid * VEC; base < N_ELEMS; base += stride_elems) {
+    const int outer = base / HIDDEN;
+    const int inner = base % HIDDEN;
+    const int xw1_off = outer * 2 * HIDDEN + inner;
+    const int xw3_off = xw1_off + HIDDEN;
+
+    float4 x1_raw = *reinterpret_cast<const float4*>(&xw13[xw1_off]);
+    float4 x3_raw = *reinterpret_cast<const float4*>(&xw13[xw3_off]);
+    float4 g_raw  = *reinterpret_cast<const float4*>(&grad_x2[base]);
+
+    const __hip_bfloat16 *x1 = reinterpret_cast<const __hip_bfloat16*>(&x1_raw);
+    const __hip_bfloat16 *x3 = reinterpret_cast<const __hip_bfloat16*>(&x3_raw);
+    const __hip_bfloat16 *gv = reinterpret_cast<const __hip_bfloat16*>(&g_raw);
+
+    __hip_fp8_storage_t fp8_1[VEC], fp8_3[VEC];
+    #pragma unroll
+    for (int i = 0; i < VEC; i++) {
+      const float f1 = static_cast<float>(x1[i]);
+      const float f3 = static_cast<float>(x3[i]);
+      const float fg = static_cast<float>(gv[i]);
+      const float sig = 1.0f / (1.0f + __expf(-f1));
+      const float silu = f1 * sig;
+      const float silu_prime = sig + silu * (1.0f - sig);
+      const float gs = fg * scale;
+      const float g1 = gs * silu_prime * f3;
+      const float g3 = gs * silu;
+      local_max = fmaxf(local_max, fmaxf(fabsf(g1), fabsf(g3)));
+      fp8_1[i] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, g1 * g_scale)), __HIP_SATFINITE, __HIP_E4M3);
+      fp8_3[i] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, g3 * g_scale)), __HIP_SATFINITE, __HIP_E4M3);
+    }
+
+    *reinterpret_cast<uint64_t*>(&grad_xw13_fp8_out[xw1_off]) = *reinterpret_cast<uint64_t*>(fp8_1);
+    *reinterpret_cast<uint64_t*>(&grad_xw13_fp8_out[xw3_off]) = *reinterpret_cast<uint64_t*>(fp8_3);
+  }
+
+  sdata[tid] = local_max;
+  __syncthreads();
+  for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+    if (tid < s) sdata[tid] = fmaxf(sdata[tid], sdata[tid + s]);
+    __syncthreads();
+  }
+  if (tid == 0) grad_amax_buf[wg] = sdata[0];
+}

tinygrad_repo/extra/llama_kernels/cast_amax/cast_amax_fwd_w13.cpp

@@ -0,0 +1,79 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+#include <hip/hip_fp8.h>
+
+#ifndef N_ELEMS
+#define N_ELEMS 234881024
+#endif
+#ifndef HIDDEN
+#define HIDDEN 14336
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % VEC == 0, "N_ELEMS must be divisible by VEC");
+static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC (so VEC loads don't straddle block boundary)");
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_silu_mul_cast_amax_w13(
+    __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, N_ELEMS
+    float*                __restrict__ amax_buf,        // fp32, NUM_WG (per-WG amaxes)
+    const __hip_bfloat16* __restrict__ xw13,            // bf16, 2*N_ELEMS
+    const float*          __restrict__ amax_state)      // fp32 scalar
+{
+  __shared__ float sdata[THREADS_PER_WG];
+
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+  const int gid = wg * THREADS_PER_WG + tid;
+  const int stride_elems = NUM_WG * THREADS_PER_WG * VEC;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  float local_max = 0.0f;
+
+  // grid-stride over 8-element groups
+  for (int base = gid * VEC; base < N_ELEMS; base += stride_elems) {
+    // interleaved xw13 layout: xw1 and xw3 are not contiguous halves
+    const int outer = base / HIDDEN;
+    const int inner = base % HIDDEN;
+    const int xw1_off = outer * 2 * HIDDEN + inner;
+    const int xw3_off = xw1_off + HIDDEN;
+
+    float4 x1_raw = *reinterpret_cast<const float4*>(&xw13[xw1_off]);
+    float4 x3_raw = *reinterpret_cast<const float4*>(&xw13[xw3_off]);
+
+    const __hip_bfloat16 *x1 = reinterpret_cast<const __hip_bfloat16*>(&x1_raw);
+    const __hip_bfloat16 *x3 = reinterpret_cast<const __hip_bfloat16*>(&x3_raw);
+
+    __hip_fp8_storage_t out[VEC];
+    #pragma unroll
+    for (int i = 0; i < VEC; i++) {
+      const float f1 = static_cast<float>(x1[i]);
+      const float f3 = static_cast<float>(x3[i]);
+      const float silu = f1 / (1.0f + __expf(-f1));
+      const float x2 = silu * f3;
+      local_max = fmaxf(local_max, fabsf(x2));
+      const float x_scaled = fmaxf(-FP8_MAX, fminf(FP8_MAX, x2 * scale));
+      out[i] = __hip_cvt_float_to_fp8(x_scaled, __HIP_SATFINITE, __HIP_E4M3);
+    }
+
+    *reinterpret_cast<uint64_t*>(&fp8_out[base]) = *reinterpret_cast<uint64_t*>(out);
+  }
+
+  // LDS tree reduction: per-workgroup amax
+  sdata[tid] = local_max;
+  __syncthreads();
+  for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+    if (tid < s) sdata[tid] = fmaxf(sdata[tid], sdata[tid + s]);
+    __syncthreads();
+  }
+
+  if (tid == 0) amax_buf[wg] = sdata[0];
+}

tinygrad_repo/extra/llama_kernels/fp8_transpose/__init__.py

@@ -0,0 +1,41 @@
+from __future__ import annotations
+import functools, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import THREADS_PER_WG, alloc_like, dname_of, compile_hip
+
+TILE = 64
+
+@functools.cache
+def _custom_fp8_transpose(out:UOp, inp:UOp, dname:str) -> UOp:
+  M, N = inp.shape
+  num_wg = (M // TILE) * (N // TILE)
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
+  mem = M * N * 2  # one byte read + one byte write per element
+  sink = UOp.sink(out.base, inp.base, threads, workgroups,
+                  arg=KernelInfo(f"fp8_transpose_{M}_{N}",
+                                 estimates=Estimates(ops=M*N, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"fp8_transpose.cpp").read_text()
+  defines = [f"-DM_DIM={M}", f"-DN_DIM={N}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def fast_fp8_transpose(t:Tensor) -> Tensor:
+  assert t.ndim == 2, f"fast_fp8_transpose needs 2D input, got shape {t.shape}"
+  assert t.dtype in dtypes.fp8s, f"fast_fp8_transpose needs fp8 dtype, got {t.dtype}"
+  M, N = t.shape
+  assert M % TILE == 0 and N % TILE == 0, f"M={M}, N={N} must be multiples of {TILE}"
+
+  device = t.device
+  axis = t.uop.axis if isinstance(device, tuple) else None
+  out_axis = None
+  if axis == 0: out_axis = 1
+  elif axis == 1: out_axis = 0
+  elif axis is not None:
+    raise ValueError(f"fast_fp8_transpose: unsupported axis {axis}")
+
+  out = alloc_like((N, M), t.dtype, device, out_axis)
+  fxn = functools.partial(_custom_fp8_transpose, dname=dname_of(device))
+  out, _ = Tensor.custom_kernel(out, t, fxn=fxn)
+  return out

tinygrad_repo/extra/llama_kernels/fp8_transpose/fp8_transpose.cpp

@@ -0,0 +1,74 @@
+#include <hip/hip_runtime.h>
+
+// LDS-staged 64x64 fp8 transpose.
+//   in : (M_DIM, N_DIM) fp8 contiguous
+//   out: (N_DIM, M_DIM) fp8 contiguous, out[c][r] = in[r][c]
+//
+// One WG processes one 64x64 output tile.  Each thread reads one uint4 (16 fp8) coalesced
+// from input rows, stages into LDS, then writes one uint4 coalesced to the output (whose
+// 16 fp8 come from 16 different input rows via in-LDS gather).
+//
+// LDS layout: lds[64][LDS_STRIDE] with LDS_STRIDE=65 (1 byte pad) to mitigate bank conflicts
+// during the column-direction read of the write phase.
+
+#ifndef M_DIM
+#define M_DIM 16384
+#endif
+#ifndef N_DIM
+#define N_DIM 28672
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int TILE = 64;
+constexpr int VEC = 16;                    // fp8 per uint4 (128-bit) load/store
+constexpr int LDS_PAD = 1;
+constexpr int LDS_STRIDE = TILE + LDS_PAD; // 65 fp8 per row
+
+static_assert(THREADS_PER_WG * VEC == TILE * TILE, "256 threads * 16 fp8 = 64*64");
+static_assert(M_DIM % TILE == 0, "M_DIM must be a multiple of 64");
+static_assert(N_DIM % TILE == 0, "N_DIM must be a multiple of 64");
+
+constexpr int N_TILES_N = N_DIM / TILE;
+
+struct alignas(16) fp8x16 { uint8_t v[16]; };
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fp8_transpose(uint8_t* __restrict__ out,         // (N_DIM, M_DIM)
+              const uint8_t* __restrict__ in)    // (M_DIM, N_DIM)
+{
+  __shared__ uint8_t lds[TILE * LDS_STRIDE];
+
+  const int tid     = threadIdx.x;
+  const int wg_id   = blockIdx.x;
+  const int tile_r  = wg_id / N_TILES_N;       // tile index along M dim of input
+  const int tile_c  = wg_id % N_TILES_N;       // tile index along N dim of input
+
+  const int a = tid / (TILE / VEC);            // 0..63 (row within tile during read; col within tile during write)
+  const int b = tid % (TILE / VEC);            // 0..3
+  const int b16 = b * VEC;                     // 0,16,32,48
+
+  // ---- Read phase: input rows -> LDS rows
+  {
+    const long long src = (long long)(tile_r * TILE + a) * (long long)N_DIM
+                        + (long long)(tile_c * TILE + b16);
+    fp8x16 v = *reinterpret_cast<const fp8x16*>(&in[src]);
+    *reinterpret_cast<fp8x16*>(&lds[a * LDS_STRIDE + b16]) = v;
+  }
+  __syncthreads();
+
+  // ---- Write phase: LDS columns (gathered) -> output rows
+  // out[(tile_c*TILE + a)][(tile_r*TILE + b16 + i)] = in[(tile_r*TILE + b16 + i)][(tile_c*TILE + a)]
+  //                                                  = lds[b16 + i][a]
+  {
+    fp8x16 v;
+    #pragma unroll
+    for (int i = 0; i < VEC; ++i) {
+      v.v[i] = lds[(b16 + i) * LDS_STRIDE + a];
+    }
+    const long long dst = (long long)(tile_c * TILE + a) * (long long)M_DIM
+                        + (long long)(tile_r * TILE + b16);
+    *reinterpret_cast<fp8x16*>(&out[dst]) = v;
+  }
+}

tinygrad_repo/extra/llama_kernels/fused_ce/__init__.py

@@ -0,0 +1,89 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+
+@functools.cache
+def _custom_fused_ce_loss_fwd(loss_out:UOp, max_out:UOp, lse_out:UOp, logits:UOp, targets:UOp,
+                              vocab:int, rows:int, label_smoothing:float) -> UOp:
+  row = UOp.range(rows, 0)
+
+  v_max = UOp.range(vocab, 1, axis_type=AxisType.REDUCE)
+  row_max = logits[row, v_max].cast(dtypes.float).reduce(v_max, arg=Ops.MAX)
+
+  v_lse = UOp.range(vocab, 2, axis_type=AxisType.REDUCE)
+  row_lse = (logits[row, v_lse].cast(dtypes.float) - row_max).exp().reduce(v_lse, arg=Ops.ADD).log() + row_max
+
+  v_smooth = UOp.range(vocab, 3, axis_type=AxisType.REDUCE)
+  target = logits[row, targets[row].cast(dtypes.weakint)].cast(dtypes.float)
+  mean_logits = logits[row, v_smooth].cast(dtypes.float).reduce(v_smooth, arg=Ops.ADD) / vocab
+  loss = row_lse - (1.0 - label_smoothing) * target - label_smoothing * mean_logits
+  stores = UOp.group(loss_out[row].store(loss), max_out[row].store(row_max), lse_out[row].store(row_lse))
+
+  return stores.end(row).sink(arg=KernelInfo(f"fused_ce_loss_fwd_{rows}_{vocab}"))
+
+@functools.cache
+def _custom_fused_ce_loss_bwd(d_logits:UOp, logits:UOp, lse:UOp, targets:UOp, scale:UOp,
+                              vocab:int, rows:int, label_smoothing:float) -> UOp:
+  row = UOp.range(rows, 0)
+  v = UOp.range(vocab, 1)
+
+  prob = (logits[row, v].cast(dtypes.float) - lse[row]).exp()
+  target = v.eq(targets[row].cast(dtypes.weakint)).where(1.0 - label_smoothing, 0.0)
+  smooth = label_smoothing / vocab
+  grad = (prob - target - smooth) * scale[0]
+
+  return d_logits[row, v].store(grad.cast(d_logits.dtype.base)).end(v, row).sink(arg=KernelInfo(f"fused_ce_loss_bwd_{rows}_{vocab}"))
+
+def _fused_ce_loss_bwd(gradient:UOp, kernel:UOp, label_smoothing:float):
+  # NOTE: forward inputs are (loss_out, max_out, lse_out, logits, targets)
+  # gradient is the upstream grad w.r.t. per-row loss (shape: (rows,) fp32)
+  _, _, lse_u, logits_u, targets_u = kernel.src[1:]
+  device = logits_u.device
+  rows, VOCAB = logits_u.shape  # (rows, VOCAB) after reshape
+  if isinstance(device, tuple):
+    axis = logits_u.axis
+    ndev = len(device)
+    d_logits = Tensor(Tensor.invalids(rows // ndev, VOCAB, dtype=dtypes.bfloat16, device=device).uop.multi(axis), device=device)
+    rows_per_dev = rows // ndev
+  else:
+    d_logits = Tensor.invalids(rows, VOCAB, dtype=dtypes.bfloat16, device=device)
+    rows_per_dev = rows
+  # NOTE: .mean() backward gives same grad per row (1/N), so broadcast is safe; take scalar
+  scale = Tensor(gradient, device=device).float().reshape(-1)[0:1].contiguous()
+  logits_t = Tensor(logits_u.after(kernel), device=device)
+  lse_t = Tensor(lse_u.after(kernel), device=device)
+  targets_t = Tensor(targets_u, device=device)
+  fxn = functools.partial(_custom_fused_ce_loss_bwd, vocab=VOCAB, rows=rows_per_dev, label_smoothing=label_smoothing)
+  d_logits, *_ = Tensor.custom_kernel(d_logits, logits_t, lse_t, targets_t, scale, fxn=fxn)
+  return (None, None, None, d_logits.uop, None)
+
+def fused_ce_loss(logits:Tensor, targets:Tensor, label_smoothing:float=0.1) -> Tensor:
+  # NOTE: fused sparse_categorical_crossentropy with label smoothing, returns mean loss scalar
+  assert logits.dtype == dtypes.bfloat16, f"expected bf16, got {logits.dtype}"
+  assert logits.ndim == 3, f"expected (MBS, SEQ, VOCAB), got {logits.shape}"
+  MBS, SEQ, VOCAB = logits.shape
+  rows = MBS * SEQ
+  if isinstance(logits.device, tuple):
+    axis = logits.uop.axis
+    assert axis in (0, 1), f"unsupported sharding axis={axis} for CE loss"
+    ndev = len(logits.device)
+    loss_out = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.multi(0),
+                      device=logits.device)
+    max_out  = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.multi(0),
+                      device=logits.device)
+    lse_out  = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.multi(0),
+                      device=logits.device)
+    rows_per_dev = rows // ndev
+  else:
+    loss_out = Tensor.invalids(rows, dtype=dtypes.float32, device=logits.device)
+    max_out  = Tensor.invalids(rows, dtype=dtypes.float32, device=logits.device)
+    lse_out  = Tensor.invalids(rows, dtype=dtypes.float32, device=logits.device)
+    rows_per_dev = rows
+  logits_flat = logits.reshape(rows, VOCAB)
+  targets_flat = targets.reshape(-1).cast(dtypes.int32)
+  fxn = functools.partial(_custom_fused_ce_loss_fwd, vocab=VOCAB, rows=rows_per_dev,
+                          label_smoothing=label_smoothing)
+  loss_out, max_out, lse_out, *_ = Tensor.custom_kernel(
+    loss_out, max_out, lse_out, logits_flat, targets_flat,
+    fxn=fxn, grad_fxn=functools.partial(_fused_ce_loss_bwd, label_smoothing=label_smoothing))
+  return loss_out.mean()

tinygrad_repo/extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/__init__.py

@@ -0,0 +1,153 @@
+from __future__ import annotations
+import functools, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, alloc_like, alloc_local, scalar_amax, dname_of, compile_hip
+
+def _src() -> str: return (pathlib.Path(__file__).parent/"fused_rmsnorm_mul_quantize_fp8.cpp").read_text()
+def _src_bwd() -> str: return (pathlib.Path(__file__).parent/"fused_rmsnorm_mul_quantize_fp8_bwd.cpp").read_text()
+
+@functools.cache
+def _custom_fwd(fp8_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
+                x:UOp, weight:UOp, amax_state:UOp, dname:str, eps_val:float) -> UOp:
+  MBS, SEQ, HIDDEN = x.shape
+  n_elems = MBS * SEQ * HIDDEN
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 + n_elems + MBS * SEQ * 4 + n_elems + HIDDEN * 2 + NUM_WG * 4 + 4
+  sink = UOp.sink(fp8_out.base, x_normed_out.base, rrms_out.base, amax_buf.base,
+                  x.base, weight.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"fused_rmsnorm_mul_quantize_fp8_{n_elems}_h{HIDDEN}_eps{eps_val:.0e}",
+                                 estimates=Estimates(ops=6*n_elems, mem=mem)))
+  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
+             f"-DEPS_LITERAL={eps_val}f"]
+  src = _src()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+@functools.cache
+def _custom_fwd_add(fp8_out:UOp, h_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
+                    x:UOp, residual:UOp, weight:UOp, amax_state:UOp, dname:str, eps_val:float) -> UOp:
+  MBS, SEQ, HIDDEN = x.shape
+  n_elems = MBS * SEQ * HIDDEN
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 * 4 + MBS * SEQ * 4 + HIDDEN * 2 + NUM_WG * 4 + 4
+  sink = UOp.sink(fp8_out.base, h_out.base, x_normed_out.base, rrms_out.base, amax_buf.base,
+                  x.base, residual.base, weight.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"fused_add_rmsnorm_mul_quantize_fp8_{n_elems}_h{HIDDEN}_eps{eps_val:.0e}",
+                                 estimates=Estimates(ops=7*n_elems, mem=mem)))
+  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
+             f"-DEPS_LITERAL={eps_val}f", f"-DHAS_RESIDUAL=1"]
+  src = _src()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+@functools.cache
+def _custom_bwd(grad_x:UOp, grad_weight_partial:UOp,
+                grad_fp8:UOp, x_normed:UOp, rrms:UOp, weight:UOp, amax_state:UOp, dname:str) -> UOp:
+  MBS, SEQ, HIDDEN = x_normed.shape
+  n_elems = MBS * SEQ * HIDDEN
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 * 3 + NUM_WG * HIDDEN * 4 + MBS * SEQ * 4 + HIDDEN * 2 + 4
+  sink = UOp.sink(grad_x.base, grad_weight_partial.base,
+                  grad_fp8.base, x_normed.base, rrms.base, weight.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"fused_rmsnorm_mul_quantize_fp8_bwd_{n_elems}_h{HIDDEN}",
+                                 estimates=Estimates(ops=8*n_elems, mem=mem)))
+  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  src = _src_bwd()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel:UOp):
+  device = x_u.device
+  MBS, SEQ, HIDDEN = x_normed_u.shape
+  axis = x_normed_u.axis if isinstance(device, tuple) else None
+  grad_x = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, device, axis)
+  grad_weight_partial = alloc_local((NUM_WG, HIDDEN), dtypes.float32, device, axis)
+  grad_h_from_fp8 = None
+  grad_weight_uop = None
+  if fp8_grad_u is not None:
+    fxn = functools.partial(_custom_bwd, dname=dname_of(device))
+    grad_x_t, grad_weight_partial_t, *_ = Tensor.custom_kernel(
+      grad_x, grad_weight_partial,
+      Tensor(fp8_grad_u, device=device).cast(dtypes.bfloat16),
+      Tensor(x_normed_u.after(kernel), device=device),
+      Tensor(rrms_u.after(kernel), device=device),
+      Tensor(weight_u, device=device),
+      Tensor(amax_state_u, device=device), fxn=fxn)
+    grad_h_from_fp8 = grad_x_t
+    grad_weight_uop = grad_weight_partial_t.sum(axis=0).cast(dtypes.bfloat16).uop
+  if h_grad_u is not None:
+    h_grad_t = Tensor(h_grad_u, device=device).cast(dtypes.bfloat16)
+    grad_total = (grad_h_from_fp8 + h_grad_t) if grad_h_from_fp8 is not None else h_grad_t
+  else:
+    grad_total = grad_h_from_fp8
+  return grad_total.uop, grad_weight_uop
+
+def _fused_bwd(gradient:UOp, kernel:UOp):
+  # NOTE: fwd inputs (fp8_out, x_normed_out, rrms_out, amax_buf, x, weight, amax_state)
+  _, x_normed_u, rrms_u, _, x_u, weight_u, amax_state_u = kernel.src[1:]
+  grad_x, grad_w = _bwd_common(gradient, None, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel)
+  return (None, None, None, None, grad_x, grad_w, None)
+
+def _fused_add_bwd(*args, **kwargs):
+  # Two invocation modes: 1 grad => positional; >1 grads => kwarg `call=`.
+  # Outputs: (fp8_out, h_out, x_normed_out, rrms_out, amax_buf). Both fp8 and h may be consumed
+  # downstream — TUPLE order in gradient.py preserves kernel-output slot order.
+  # Don't dispatch by dtype: matmul's bwd emits fp8 grad as bf16 (no explicit cast), so
+  # dtype-detection collapses both into h_grad and silently drops the rmsnorm-bwd path.
+  if 'call' in kwargs:
+    kernel, all_grads = kwargs['call'], list(args)
+  else:
+    gradient, kernel = args
+    all_grads = [gradient]
+  fp8_grad_u = h_grad_u = None
+  if len(all_grads) >= 2:
+    fp8_grad_u, h_grad_u = all_grads[0], all_grads[1]
+  elif len(all_grads) == 1:
+    g = all_grads[0]
+    if g.dtype == dtypes.bfloat16: h_grad_u = g
+    else: fp8_grad_u = g
+  _, _, x_normed_u, rrms_u, _, x_u, _, weight_u, amax_state_u = kernel.src[1:]
+  grad_h, grad_w = _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel)
+  return (None, None, None, None, None, grad_h, grad_h, grad_w, None)
+
+def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor]:
+  # NOTE: rmsnorm(x) * weight -> fp8 + amax. Returns (fp8, inv_scale, new_amax, x_normed, rrms).
+  # x_normed + rrms are saved for the rmsnorm backward (also recomputed here from x regs).
+  assert x.dtype == dtypes.bfloat16 and weight.dtype == dtypes.bfloat16
+  assert x.shape[-1] == weight.shape[-1], f"HIDDEN mismatch: x={x.shape}, weight={weight.shape}"
+  MBS, SEQ, HIDDEN = x.shape
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  if isinstance(x.device, tuple): assert axis in (None, 0, 1), f"unsupported sharding axis={axis}"
+  fp8_out      = alloc_like((MBS, SEQ, HIDDEN), fp8_dtype,       x.device, axis)
+  x_normed_out = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, x.device, axis)
+  rrms_out     = alloc_like((MBS, SEQ),         dtypes.float32,  x.device, axis)
+  amax_buf     = alloc_local((NUM_WG,),         dtypes.float32,  x.device, axis)
+  fxn = functools.partial(_custom_fwd, dname=dname_of(x.device), eps_val=eps)
+  fp8_out, x_normed_out, rrms_out, amax_buf, *_ = Tensor.custom_kernel(
+    fp8_out, x_normed_out, rrms_out, amax_buf, x, weight, amax_state, fxn=fxn, grad_fxn=_fused_bwd)
+  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
+  return fp8_out, inv_scale, scalar_amax(amax_buf), x_normed_out, rrms_out
+
+def fused_add_rmsnorm_mul_quantize_fp8(x:Tensor, residual:Tensor, weight:Tensor, amax_state:Tensor,
+                                       eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+  # NOTE: h = x + residual; y_normed = rmsnorm(h); fp8 = quantize(y_normed * weight).
+  # Returns (fp8, inv_scale, new_amax, h, x_normed, rrms). h is also written so downstream can
+  # reuse it without recomputing x+residual — eliminates the separate residual-add kernel.
+  assert x.dtype == dtypes.bfloat16 and residual.dtype == dtypes.bfloat16 and weight.dtype == dtypes.bfloat16
+  assert x.shape == residual.shape
+  MBS, SEQ, HIDDEN = x.shape
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  if isinstance(x.device, tuple): assert axis in (None, 0, 1), f"unsupported sharding axis={axis}"
+  fp8_out      = alloc_like((MBS, SEQ, HIDDEN), fp8_dtype,       x.device, axis)
+  h_out        = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, x.device, axis)
+  x_normed_out = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, x.device, axis)
+  rrms_out     = alloc_like((MBS, SEQ),         dtypes.float32,  x.device, axis)
+  amax_buf     = alloc_local((NUM_WG,),         dtypes.float32,  x.device, axis)
+  fxn = functools.partial(_custom_fwd_add, dname=dname_of(x.device), eps_val=eps)
+  fp8_out, h_out, x_normed_out, rrms_out, amax_buf, *_ = Tensor.custom_kernel(
+    fp8_out, h_out, x_normed_out, rrms_out, amax_buf, x, residual, weight, amax_state,
+    fxn=fxn, grad_fxn=_fused_add_bwd)
+  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
+  return fp8_out, inv_scale, scalar_amax(amax_buf), h_out, x_normed_out, rrms_out

tinygrad_repo/extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/fused_rmsnorm_mul_quantize_fp8.cpp

@@ -0,0 +1,155 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+#include <hip/hip_fp8.h>
+
+// Fuses the full pre-matmul preparation for a layer into a single HBM pass:
+//   y = rmsnorm(x) * weight            (reduce-mean-square + rsqrt + per-elem mul)
+//   fp8 = fp8_sat(y * (FP8_MAX / amax_state))
+// Also writes:
+//   rrms[row]        — saved for the rmsnorm backward
+//   amax_buf[wg]     — per-WG |y| partials, reduced later to update amax_state
+//
+// Layout: one WG per row, ROWS_PER_WG rows per WG via grid-stride (ROWS = N_ELEMS / HIDDEN).
+// Each thread handles HIDDEN / THREADS_PER_WG elements per row.
+
+#ifndef N_ELEMS
+#define N_ELEMS 67108864
+#endif
+#ifndef HIDDEN
+#define HIDDEN 4096
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+#ifndef EPS_LITERAL
+#define EPS_LITERAL 1e-5f
+#endif
+#ifndef HAS_RESIDUAL
+#define HAS_RESIDUAL 0
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % HIDDEN == 0, "N_ELEMS must be a multiple of HIDDEN");
+static_assert(HIDDEN % (THREADS_PER_WG * VEC) == 0, "HIDDEN must be divisible by THREADS_PER_WG*VEC");
+
+constexpr int ROWS = N_ELEMS / HIDDEN;
+constexpr int ELEMS_PER_THREAD = HIDDEN / THREADS_PER_WG;  // each thread sees this many elems per row
+constexpr int VECS_PER_THREAD = ELEMS_PER_THREAD / VEC;    // number of 8-wide vec loads
+
+#if HAS_RESIDUAL
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_add_rmsnorm_mul_quantize_fp8(
+    __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, ROWS*HIDDEN
+    __hip_bfloat16*       __restrict__ h_out,           // bf16, ROWS*HIDDEN — x + residual (saved for downstream)
+    __hip_bfloat16*       __restrict__ x_normed_out,    // bf16, ROWS*HIDDEN
+    float*                __restrict__ rrms_out,        // fp32, ROWS
+    float*                __restrict__ amax_buf,        // fp32, NUM_WG
+    const __hip_bfloat16* __restrict__ x,               // bf16, ROWS*HIDDEN
+    const __hip_bfloat16* __restrict__ residual,        // bf16, ROWS*HIDDEN — added into x before rmsnorm
+    const __hip_bfloat16* __restrict__ weight,          // bf16, HIDDEN
+    const float*          __restrict__ amax_state)      // fp32 scalar
+{
+#else
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_rmsnorm_mul_quantize_fp8(
+    __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, ROWS*HIDDEN
+    __hip_bfloat16*       __restrict__ x_normed_out,    // bf16, ROWS*HIDDEN (saved for rmsnorm bwd)
+    float*                __restrict__ rrms_out,        // fp32, ROWS (fp32 to match rmsnorm_bwd.cpp expectation)
+    float*                __restrict__ amax_buf,        // fp32, NUM_WG per-WG partials
+    const __hip_bfloat16* __restrict__ x,               // bf16, ROWS*HIDDEN
+    const __hip_bfloat16* __restrict__ weight,          // bf16, HIDDEN (per-hidden scale)
+    const float*          __restrict__ amax_state)      // fp32 scalar
+{
+#endif
+  __shared__ float sdata[THREADS_PER_WG];
+
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  const float inv_hidden = 1.0f / static_cast<float>(HIDDEN);
+  float local_max = 0.0f;
+
+  // Grid-stride over rows. Each WG processes rows (wg, wg+NUM_WG, wg+2*NUM_WG, ...).
+  for (int row = wg; row < ROWS; row += NUM_WG) {
+    const int row_off = row * HIDDEN;
+
+    // Load row (+ residual if present) into registers.
+    float regs[ELEMS_PER_THREAD];
+    float sum_sq = 0.0f;
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      float4 raw = *reinterpret_cast<const float4*>(&x[row_off + h_base]);
+      const __hip_bfloat16 *xi = reinterpret_cast<const __hip_bfloat16*>(&raw);
+#if HAS_RESIDUAL
+      float4 res_raw = *reinterpret_cast<const float4*>(&residual[row_off + h_base]);
+      const __hip_bfloat16 *ri = reinterpret_cast<const __hip_bfloat16*>(&res_raw);
+      __hip_bfloat16 h_buf[VEC];
+#endif
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+#if HAS_RESIDUAL
+        const float f = static_cast<float>(xi[i]) + static_cast<float>(ri[i]);
+        h_buf[i] = static_cast<__hip_bfloat16>(f);
+#else
+        const float f = static_cast<float>(xi[i]);
+#endif
+        regs[v * VEC + i] = f;
+        sum_sq += f * f;
+      }
+#if HAS_RESIDUAL
+      *reinterpret_cast<float4*>(&h_out[row_off + h_base]) = *reinterpret_cast<float4*>(h_buf);
+#endif
+    }
+
+    // LDS tree-reduce sum_sq across the WG.
+    sdata[tid] = sum_sq;
+    __syncthreads();
+    for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+      if (tid < s) sdata[tid] = sdata[tid] + sdata[tid + s];
+      __syncthreads();
+    }
+    const float mean_sq = sdata[0] * inv_hidden;
+    const float rrms = 1.0f / sqrtf(mean_sq + EPS_LITERAL);
+
+    if (tid == 0) rrms_out[row] = rrms;
+
+    // Normalize, multiply by weight, quantize. Also write x_normed (for rmsnorm bwd).
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      float4 w_raw = *reinterpret_cast<const float4*>(&weight[h_base]);
+      const __hip_bfloat16 *wi = reinterpret_cast<const __hip_bfloat16*>(&w_raw);
+
+      __hip_fp8_storage_t out[VEC];
+      __hip_bfloat16 xn[VEC];
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+        const float x_normed = regs[v * VEC + i] * rrms;
+        xn[i] = static_cast<__hip_bfloat16>(x_normed);
+        const float y = x_normed * static_cast<float>(wi[i]);
+        local_max = fmaxf(local_max, fabsf(y));
+        const float scaled = fmaxf(-FP8_MAX, fminf(FP8_MAX, y * scale));
+        out[i] = __hip_cvt_float_to_fp8(scaled, __HIP_SATFINITE, __HIP_E4M3);
+      }
+      *reinterpret_cast<uint64_t*>(&fp8_out[row_off + h_base]) = *reinterpret_cast<uint64_t*>(out);
+      *reinterpret_cast<float4*>(&x_normed_out[row_off + h_base]) = *reinterpret_cast<float4*>(xn);
+    }
+    __syncthreads();  // before next row's sum_sq reduce reuses sdata
+  }
+
+  // Final per-WG amax reduce.
+  sdata[tid] = local_max;
+  __syncthreads();
+  for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+    if (tid < s) sdata[tid] = fmaxf(sdata[tid], sdata[tid + s]);
+    __syncthreads();
+  }
+  if (tid == 0) amax_buf[wg] = sdata[0];
+}

tinygrad_repo/extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/fused_rmsnorm_mul_quantize_fp8_bwd.cpp

@@ -0,0 +1,147 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+
+// Full backward for fused_rmsnorm_mul_quantize_fp8.cpp. One HBM pass per row produces:
+//   grad_x (bf16)              — gradient w.r.t. pre-rmsnorm x
+//   grad_weight_partial (fp32) — per-WG partial of the weight gradient, reduced later
+//
+// Input (all read):
+//   grad_fp8 (bf16)     — upstream grad w.r.t. fp8_out (bf16-typed gradient value)
+//   x_normed (bf16)     — saved from the fwd kernel, shape (ROWS, HIDDEN)
+//   rrms (fp32)         — saved rrms per row
+//   weight (bf16)       — per-HIDDEN rmsnorm weight
+//   amax_state (bf16)   — delayed amax used to compute the fp8 scale in fwd
+//
+// Chain: y = x_normed * weight; fp8 = sat(y * scale). Through STE: grad_y = grad_fp8 * scale.
+//        grad_x_normed = grad_y * weight.
+//        grad_weight   = sum_rows(grad_y * x_normed).
+//        grad_x        = rrms * (grad_x_normed - x_normed * mean(grad_x_normed * x_normed, last_dim)).
+
+#ifndef N_ELEMS
+#define N_ELEMS 67108864
+#endif
+#ifndef HIDDEN
+#define HIDDEN 4096
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % HIDDEN == 0, "N_ELEMS must be a multiple of HIDDEN");
+static_assert(HIDDEN % (THREADS_PER_WG * VEC) == 0, "HIDDEN must be divisible by THREADS_PER_WG*VEC");
+
+constexpr int ROWS = N_ELEMS / HIDDEN;
+constexpr int ELEMS_PER_THREAD = HIDDEN / THREADS_PER_WG;
+constexpr int VECS_PER_THREAD = ELEMS_PER_THREAD / VEC;
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_rmsnorm_mul_quantize_fp8_bwd(
+    __hip_bfloat16*       __restrict__ grad_x,              // out: bf16, ROWS*HIDDEN
+    float*                __restrict__ grad_weight_partial, // out: fp32, NUM_WG*HIDDEN
+    const __hip_bfloat16* __restrict__ grad_fp8,            // in:  bf16, ROWS*HIDDEN (grad of fp8_out)
+    const __hip_bfloat16* __restrict__ x_normed,            // in:  bf16, ROWS*HIDDEN
+    const float*          __restrict__ rrms,                // in:  fp32, ROWS
+    const __hip_bfloat16* __restrict__ weight,              // in:  bf16, HIDDEN
+    const float*          __restrict__ amax_state)          // in:  fp32 scalar
+{
+  __shared__ float sdata[THREADS_PER_WG];
+
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  const float inv_hidden = 1.0f / static_cast<float>(HIDDEN);
+
+  // Per-thread accumulator for grad_weight (across all rows this WG touches).
+  float gw_accum[ELEMS_PER_THREAD];
+  #pragma unroll
+  for (int i = 0; i < ELEMS_PER_THREAD; i++) gw_accum[i] = 0.0f;
+
+  // Preload weight into registers (same across rows). Use ELEMS_PER_THREAD entries.
+  float w_regs[ELEMS_PER_THREAD];
+  #pragma unroll
+  for (int v = 0; v < VECS_PER_THREAD; v++) {
+    const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+    float4 w_raw = *reinterpret_cast<const float4*>(&weight[h_base]);
+    const __hip_bfloat16 *wi = reinterpret_cast<const __hip_bfloat16*>(&w_raw);
+    #pragma unroll
+    for (int i = 0; i < VEC; i++) w_regs[v * VEC + i] = static_cast<float>(wi[i]);
+  }
+
+  for (int row = wg; row < ROWS; row += NUM_WG) {
+    const int row_off = row * HIDDEN;
+    const float rrms_v = rrms[row];
+
+    // Load grad_fp8 and x_normed rows into registers, compute grad_y and grad_x_normed.
+    float g_y_regs[ELEMS_PER_THREAD];
+    float xn_regs[ELEMS_PER_THREAD];
+    float g_xn_regs[ELEMS_PER_THREAD];  // grad_x_normed
+    float local_dot = 0.0f;  // sum(grad_x_normed * x_normed) for mean
+
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      float4 g_raw  = *reinterpret_cast<const float4*>(&grad_fp8[row_off + h_base]);
+      float4 xn_raw = *reinterpret_cast<const float4*>(&x_normed[row_off + h_base]);
+      const __hip_bfloat16 *gi  = reinterpret_cast<const __hip_bfloat16*>(&g_raw);
+      const __hip_bfloat16 *xni = reinterpret_cast<const __hip_bfloat16*>(&xn_raw);
+
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+        const int idx = v * VEC + i;
+        const float g_y = static_cast<float>(gi[i]) * scale;
+        const float xn  = static_cast<float>(xni[i]);
+        g_y_regs[idx] = g_y;
+        xn_regs[idx]  = xn;
+        g_xn_regs[idx] = g_y * w_regs[idx];           // grad_x_normed = grad_y * weight
+        gw_accum[idx] += g_y * xn;                    // grad_weight contrib
+        local_dot += g_xn_regs[idx] * xn;             // for mean
+      }
+    }
+
+    // LDS reduce local_dot to sdata[0].
+    sdata[tid] = local_dot;
+    __syncthreads();
+    for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+      if (tid < s) sdata[tid] = sdata[tid] + sdata[tid + s];
+      __syncthreads();
+    }
+    const float mean_term = sdata[0] * inv_hidden;
+
+    // Compute grad_x = rrms * (grad_x_normed - x_normed * mean_term) and write.
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      __hip_bfloat16 out[VEC];
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+        const int idx = v * VEC + i;
+        const float dx = rrms_v * (g_xn_regs[idx] - xn_regs[idx] * mean_term);
+        out[i] = static_cast<__hip_bfloat16>(dx);
+      }
+      *reinterpret_cast<float4*>(&grad_x[row_off + h_base]) = *reinterpret_cast<float4*>(out);
+    }
+    __syncthreads();
+  }
+
+  // Write this WG's grad_weight partial to HBM (fp32, NUM_WG x HIDDEN layout).
+  const int gw_row_off = wg * HIDDEN;
+  #pragma unroll
+  for (int v = 0; v < VECS_PER_THREAD; v++) {
+    const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+    // Write 8 fp32 values with two float4 stores.
+    float4 out_lo, out_hi;
+    out_lo.x = gw_accum[v * VEC + 0]; out_lo.y = gw_accum[v * VEC + 1];
+    out_lo.z = gw_accum[v * VEC + 2]; out_lo.w = gw_accum[v * VEC + 3];
+    out_hi.x = gw_accum[v * VEC + 4]; out_hi.y = gw_accum[v * VEC + 5];
+    out_hi.z = gw_accum[v * VEC + 6]; out_hi.w = gw_accum[v * VEC + 7];
+    *reinterpret_cast<float4*>(&grad_weight_partial[gw_row_off + h_base + 0]) = out_lo;
+    *reinterpret_cast<float4*>(&grad_weight_partial[gw_row_off + h_base + 4]) = out_hi;
+  }
+}

tinygrad_repo/extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -0,0 +1,98 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.dtype import AddrSpace
+from tinygrad.helpers import prod
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, alloc_like, alloc_local, scalar_amax
+
+@functools.cache
+def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_state:UOp) -> UOp:
+  VEC = 8
+  n_elems = prod(x.shape)
+  assert n_elems % (NUM_WG * THREADS_PER_WG * VEC) == 0
+  assert amax_partial.shape[0] == NUM_WG
+
+  x = x.reshape(n_elems)
+  fp8_out = fp8_out.reshape(n_elems)
+
+  wg = UOp.range(NUM_WG, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  it = UOp.range((n_elems // VEC) // (NUM_WG * THREADS_PER_WG), 2, AxisType.LOOP)
+  lane = UOp.range(VEC, 3, AxisType.UNROLL)
+
+  idx = (((it * NUM_WG + wg) * THREADS_PER_WG + tid) * VEC) + lane
+
+  scale = FP8_MAX / (amax_state[0].cast(dtypes.float) + 1e-8)
+  x_f = x[idx].cast(dtypes.float)
+  abs_x = (x_f < 0.0).where(-x_f, x_f)
+  scaled = (x_f * scale).maximum(-FP8_MAX).minimum(FP8_MAX)
+
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  lane_max = abs_x.reduce(lane, arg=Ops.MAX)
+
+  lmax = UOp.placeholder((1,), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  lmax_init = lmax.after(wg, tid)[0].store(0.0)
+  lmax_prev = lmax.after(lmax_init, it)[0]
+  lmax_store = lmax.after(fp8_store)[0].store(lmax_prev.maximum(lane_max))
+  lmax_val = lmax.after(lmax_store.end(it))[0]
+
+  lds = UOp.placeholder((THREADS_PER_WG,), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
+  lds = lds.after(lds[tid].store(lmax_val).barrier())
+
+  step = THREADS_PER_WG // 2
+  while step:
+    active = tid < step
+    other = lds[tid + step].load(UOp.const(dtypes.float, 0.0), active)
+    lds = lds.after(lds[tid].store(lds[tid].maximum(other), gate=active).barrier())
+    step //= 2
+
+  amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])
+  return amax_store.end(tid, wg).sink(arg=KernelInfo(f"quantize_fp8_with_amax_{n_elems}", opts_to_apply=()))
+
+@functools.cache
+def _custom_quantize_fp8_scalar(fp8_out:UOp, x:UOp, amax_state:UOp) -> UOp:
+  n_elems = prod(x.shape)
+  i = UOp.range(n_elems, 0)
+
+  x_f = x.reshape(n_elems)[i].cast(dtypes.float)
+  scale = FP8_MAX / (amax_state[0].cast(dtypes.float) + 1e-8)
+  store = fp8_out.reshape(n_elems)[i].store((x_f * scale).cast(fp8_out.dtype.base))
+
+  return store.end(i).sink(arg=KernelInfo(f"quantize_fp8_scalar_{n_elems}"))
+
+def _quantize_fp8_delayed_bwd(gradient:UOp, kernel:UOp):
+  # NOTE: STE-equivalent backward — grad_x = grad_fp8 * scale, scale = FP8_MAX / amax_state.
+  # `gradient` is bf16 grad w.r.t. fp8 output (asm_gemm bwd already applied x_scale).
+  _, _, x, amax_state = kernel.src[1:]
+  device = x.device
+  scale = FP8_MAX / (Tensor(amax_state, device=device).float() + 1e-8)
+  grad_x = (Tensor(gradient, device=device).float() * scale).cast(dtypes.bfloat16)
+  return (None, None, grad_x.uop, None)
+
+def quantize_fp8_delayed(x:Tensor, amax_state:Tensor, fp8_dtype=dtypes.fp8e4m3) -> tuple[Tensor, Tensor, Tensor, UOp]:
+  # NOTE: one-pass bf16 -> fp8 quantize with delayed scaling. Returns (fp8, inv_scale, new_amax, store_effect).
+  # Fused kernel reads x once and writes fp8 + per-WG |x| partials (then a small reduce produces scalar new_amax).
+  # store_effect writes new_amax into amax_state's buffer — the caller must thread it into a realized
+  # output via `.after(store_effect)`. Calling `amax_state.assign(new_amax)` inside a grad_fxn does
+  # NOT work because .assign mutates only the temp Tensor's .uop, not the original layer-owned buffer.
+  assert x.dtype == dtypes.bfloat16, f"expected bf16, got {x.dtype}"
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  fp8_out      = alloc_like(x.shape,  fp8_dtype,      x.device, axis)
+  n_elems = prod(x.uop.shard_shape)
+  assert n_elems % NUM_WG == 0, f"{n_elems=} must divide over {NUM_WG=}"
+  amax_partial = alloc_local((NUM_WG,), dtypes.float32, x.device, axis)
+  fxn = _custom_quantize_fp8_with_amax
+  fp8_out, amax_partial, *_ = Tensor.custom_kernel(fp8_out, amax_partial, x, amax_state,
+                                                    fxn=fxn, grad_fxn=_quantize_fp8_delayed_bwd)
+  new_amax = scalar_amax(amax_partial)
+  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
+  store_effect = amax_state.uop.store(new_amax.uop)
+  return fp8_out, inv_scale, new_amax, store_effect
+
+def quantize_fp8_scalar(x:Tensor, amax_state:Tensor, fp8_dtype=dtypes.fp8e4m3) -> Tensor:
+  # NOTE: pure one-pass bf16 -> fp8 quantize with delayed scalar scale. No amax computation.
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  fp8_out = alloc_like(x.shape, fp8_dtype, x.device, axis)
+  fxn = _custom_quantize_fp8_scalar
+  fp8_out, *_ = Tensor.custom_kernel(fp8_out, x, amax_state, fxn=fxn)
+  return fp8_out

tinygrad_repo/extra/llama_kernels/rmsnorm/__init__.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+import functools
+from tinygrad import Tensor
+from tinygrad.uop.ops import UOp
+
+def rmsnorm_fwd(x_in:Tensor, eps:float) -> tuple[Tensor, Tensor]:
+  x = x_in.float()
+  rrms = (x.square().mean(-1, keepdim=True) + eps).rsqrt()
+  return (x * rrms).cast(x_in.dtype), rrms
+
+@functools.cache
+def _rmsnorm_fwd_fxn(x_in_p, eps, device):
+  return rmsnorm_fwd(Tensor(x_in_p, device=device), eps)
+
+def _rmsnorm_bwd(grad:UOp, call:UOp) -> tuple:
+  x_normed = Tensor(call.gettuple(0)).float()
+  do_float = Tensor(grad).float()
+  d_x = Tensor(call.gettuple(1)) * (do_float - x_normed * (do_float * x_normed).mean(-1, keepdim=True))
+  return (d_x.cast(call.src[1].dtype).uop,)
+
+def rmsnorm(x_in:Tensor, eps:float) -> tuple[Tensor, Tensor]:
+  fxn = _rmsnorm_fwd_fxn(x_in.as_param(0).uop, eps, x_in.device)
+  call = UOp.maketuple(fxn[0].uop, fxn[1].uop).call(x_in.uop, grad_fxn=_rmsnorm_bwd)
+  return Tensor(call.gettuple(0)), Tensor(call.gettuple(1))