metal kernels

mlx/backend/cpu/unary_ops.h

@@ -120,7 +120,7 @@ Simd<uint32_t, N> fp32_to_bits(Simd<float, N> x) {
 struct ToFP8 {
   template <typename T, int N>
   Simd<uint8_t, N> operator()(Simd<T, N> f) {
-    uint32_t fp8_max = 1087 << 20;
+    uint32_t fp8_max = 543 << 21;
     auto denorm_mask = Simd<uint32_t, N>(141 << 23);
     Simd<uint32_t, N> f_bits;
     Simd<float, N> f32 = f;

mlx/backend/cuda/quantized/cuda_fp4.h

@@ -1,7 +1,22 @@
 #pragma once
 
 struct __nv_fp8_e8m0 {
-  __device__ __nv_fp8_e8m0(uint8_t x) : __x(x) {}
+  __device__ __nv_fp8_e8m0(float x) {
+    if (!std::isfinite(x)) {
+      __x = 0xFF;
+      return;
+    }
+    if (x < 0.0f) {
+      __x = 0x00;
+      return;
+    }
+    float le = std::log2f(x);
+    int n = static_cast<int>(std::nearbyintf(le));
+
+    n = n < -127 ? -127 : n;
+    n = n > 127 ? 127 : n;
+    __x = static_cast<uint8_t>(n + 127);
+  }
 
   __device__ operator float() {
     if (__x == 0xFF) {

mlx/backend/cuda/quantized/fp_quantize.cu

@@ -49,13 +49,12 @@ fp_quantize(const T* w, uint8_t* out, uint8_t* scales, size_t size) {
 
   auto grid_dim_x =
       cg::this_grid().dim_blocks().x * cg::this_grid().block_index().x;
-  size_t out_index = tidx + grid_dim_x * size_t(tidy);
+  size_t index = tidx + grid_dim_x * size_t(tidy);
-  size_t in_index = out_index;
+  if (index >= size) {
-  if (in_index >= size) {
     return;
   }
 
-  float w_thread = w[in_index];
+  float w_thread = w[index];
 
   cg::greater<float> max_op;
   auto warp = cg::tiled_partition<group_size>(cg::this_thread_block());
@@ -70,21 +69,19 @@ fp_quantize(const T* w, uint8_t* out, uint8_t* scales, size_t size) {
   scale = float(s);
 
   // Write out the scales
-  size_t gindex = in_index / group_size;
+  size_t gindex = index / group_size;
-  if (in_index % group_size == 0) {
+  if (index % group_size == 0) {
     scales[gindex] = q_scale;
   }
 
-  uint8_t output = 0;
+  uint8_t output = Quantize<bits>{}(scale == 0 ? 0.0f : w_thread / scale);
-  uint8_t val = Quantize<bits>{}(scale == 0 ? 0.0f : w_thread / scale);
-  output = val;
   if (bits == 4) {
-    uint8_t sval = warp.shfl_down(val, 1);
+    uint8_t sval = warp.shfl_down(output, 1);
     output |= sval << bits;
   }
   constexpr int pack_factor = bits == 8 ? 1 : 2;
-  if (out_index % pack_factor == 0) {
+  if (index % pack_factor == 0) {
-    out[out_index / pack_factor] = output;
+    out[index / pack_factor] = output;
   }
 }
 

mlx/backend/metal/CMakeLists.txt

@@ -29,7 +29,7 @@ make_jit_source(
   kernels/bf16_math.h
   kernels/complex.h
   kernels/defines.h)
-make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h)
+make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h kernels/fp8.h)
 make_jit_source(binary_ops)
 make_jit_source(ternary_ops)
 make_jit_source(reduce_utils kernels/atomic.h kernels/reduction/ops.h)

mlx/backend/metal/kernels/CMakeLists.txt

@@ -6,6 +6,7 @@ set(BASE_HEADERS
     defines.h
     erf.h
     expm1f.h
+    fp8.h
     utils.h)
 
 function(build_kernel_base TARGET SRCFILE DEPS)
@@ -109,7 +110,7 @@ if(NOT MLX_METAL_JIT)
     reduction/reduce_col.h
     reduction/reduce_row.h)
   build_kernel(quantized quantized.h quantized_utils.h ${STEEL_HEADERS})
-  build_kernel(fp4_quantized fp4_quantized.h quantized_utils.h ${STEEL_HEADERS})
+  build_kernel(fp_quantized fp_quantized.h quantized_utils.h ${STEEL_HEADERS})
   build_kernel(scan scan.h)
   build_kernel(softmax softmax.h)
   build_kernel(logsumexp logsumexp.h)

mlx/backend/metal/kernels/fp4.h

@@ -0,0 +1,56 @@
+#pragma once
+
+constexpr constant static float FP4_LUT[16] = {
+    +0.0f,
+    +0.5f,
+    +1.0f,
+    +1.5f,
+    +2.0f,
+    +3.0f,
+    +4.0f,
+    +6.0f,
+    -0.0f,
+    -0.5f,
+    -1.0f,
+    -1.5f,
+    -2.0f,
+    -3.0f,
+    -4.0f,
+    -6.0f};
+
+struct fp4_e2m1 {
+  fp4_e2m1(float x) {
+    if (metal::isnan(x)) {
+      bits = 0x7;
+      return;
+    }
+
+    const uint8_t sign_bit = (metal::signbit(x)) ? 0x8 : 0x0;
+    x = metal::abs(x);
+
+    if (x > 5.0f) {
+      bits = 0x7;
+    } else if (x >= 3.5f) {
+      bits = 0x6;
+    } else if (x > 2.5f) {
+      bits = 0x5;
+    } else if (x >= 1.75f) {
+      bits = 0x4;
+    } else if (x > 1.25f) {
+      bits = 0x3;
+    } else if (x >= 0.75f) {
+      bits = 0x2;
+    } else if (x > 0.25f) {
+      bits = 0x1;
+    } else {
+      bits = 0x0;
+    }
+    bits |= sign_bit;
+  }
+
+  operator float() {
+    return FP4_LUT[bits];
+  }
+
+  uint8_t bits;
+};

mlx/backend/metal/kernels/fp8.h

@@ -0,0 +1,88 @@
+#pragma once
+
+inline float fp32_from_bits(uint32_t bits) {
+  return *(reinterpret_cast<thread float*>(&bits));
+}
+inline float fp32_to_bits(float x) {
+  return *(reinterpret_cast<thread uint32_t*>(&x));
+}
+
+struct fp8_e4m3 {
+  template <typename T>
+  fp8_e4m3(T f) {
+    // From PyTorch
+    // https://github.com/pytorch/pytorch/blob/e3643e1e0e923f0fc063dfab6f45c956d568919d/c10/util/Float8_e4m3fn.h#L148
+    uint32_t fp8_max = 543 << 21;
+    uint32_t denorm_mask = 141 << 23;
+    uint32_t f_bits = fp32_to_bits(static_cast<float>(f));
+    uint32_t sign = f_bits & 0x80000000;
+    f_bits ^= sign;
+    if (f_bits >= fp8_max) {
+      // Default behavior saturates to min/max
+      bits = 0x7E;
+    } else {
+      if (f_bits < (121 << 23)) {
+        f_bits =
+            fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask));
+        bits = static_cast<uint8_t>(f_bits - denorm_mask);
+      } else {
+        // resulting mantissa is odd
+        uint8_t mant_odd = (f_bits >> 20) & 1;
+        f_bits += ((uint32_t)(7 - 127) << 23) + 0x7FFFF;
+        f_bits += mant_odd;
+        bits = static_cast<uint8_t>(f_bits >> 20);
+      }
+    }
+    bits |= static_cast<uint8_t>(sign >> 24);
+  }
+
+  operator float() {
+    // From PyTorch:
+    // https://github.com/pytorch/pytorch/blob/e3643e1e0e923f0fc063dfab6f45c956d568919d/c10/util/Float8_e4m3fn.h#L46
+    uint32_t w = static_cast<uint32_t>(bits) << 24;
+    uint32_t sign = w & 0x80000000;
+    uint32_t nonsign = w & 0x7FFFFFFF;
+
+    uint32_t renorm_shift = metal::clz(nonsign);
+    renorm_shift = renorm_shift > 4 ? renorm_shift - 4 : 0;
+
+    int32_t inf_nan_mask =
+        (static_cast<int32_t>(nonsign + 0x01000000) >> 8) & 0x7F800000;
+    int32_t zero_mask = static_cast<int32_t>(nonsign - 1) >> 31;
+    uint32_t result = sign |
+        ((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) |
+          inf_nan_mask) &
+         ~zero_mask);
+    return fp32_from_bits(result);
+  }
+
+  uint8_t bits;
+};
+
+struct fp8_e8m0 {
+  fp8_e8m0(float x) {
+    if (!metal::isfinite(x)) {
+      bits = 0xFF;
+      return;
+    }
+    if (x < 0.0f) {
+      bits = 0x00;
+      return;
+    }
+    float le = metal::log2(x);
+    int n = int(metal::round(le));
+
+    n = n < -127 ? -127 : n;
+    n = n > 127 ? 127 : n;
+    bits = static_cast<uint8_t>(n + 127);
+  }
+
+  operator float() {
+    if (bits == 0xFF) {
+      return metal::numeric_limits<float>::quiet_NaN();
+    }
+    return metal::ldexp(1.0f, static_cast<int>(bits) - 127);
+  }
+
+  uint8_t bits;
+};

mlx/backend/metal/kernels/fp_quantized.h

@@ -59,28 +59,10 @@ inline void load_vector_safe(const device T* x, thread U* x_thread, int N) {
   }
 }
 
-constexpr constant static float MXFP4_LUT[16] = {
-    +0.0f,
-    +0.5f,
-    +1.0f,
-    +1.5f,
-    +2.0f,
-    +3.0f,
-    +4.0f,
-    +6.0f,
-    -0.0f,
-    -0.5f,
-    -1.0f,
-    -1.5f,
-    -2.0f,
-    -3.0f,
-    -4.0f,
-    -6.0f};
-
 template <typename T>
 void load_mxfp4_lut(threadgroup T* lut, uint simd_gid, uint simd_lid) {
   if (simd_gid == 0 && simd_lid < 16) {
-    lut[simd_lid] = static_cast<T>(MXFP4_LUT[simd_lid]);
+    lut[simd_lid] = static_cast<T>(FP4_LUT[simd_lid]);
   }
   threadgroup_barrier(mem_flags::mem_threadgroup);
 }
@@ -1789,3 +1771,100 @@ template <
     }
   }
 }
+
+template <int bits>
+struct Quantize {
+  uint8_t operator()(float x) {
+    if constexpr (bits == 8) {
+      return fp8_e4m3(x).bits;
+    } else {
+      return fp4_e2m1(x).bits;
+    }
+  }
+};
+
+template <int bits>
+struct Dequantize {
+  float operator()(uint8_t x) {
+    if constexpr (bits == 8) {
+      return float(*(thread fp8_e4m3*)(&x));
+    } else {
+      return float(*(thread fp4_e2m1*)(&x));
+    }
+  }
+};
+
+template <typename T, const int group_size, const int bits>
+[[kernel]] void fp_quantize(
+    const device T* w [[buffer(0)]],
+    device uint8_t* out [[buffer(1)]],
+    device uint8_t* scales [[buffer(2)]],
+    uint2 tidx [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]) {
+  constexpr bool use_mx_scale = group_size == 32;
+  size_t index = tidx.x + grid_dim.x * size_t(tidx.y);
+
+  float scale;
+  float w_thread = w[index];
+  if (use_mx_scale) {
+    scale = simd_max(abs(w_thread));
+  } else {
+    float w_max_l = simd_max(tidx.x < 16 ? abs(w_thread) : 0.0);
+    float w_max_r = simd_max(tidx.x >= 16 ? abs(w_thread) : 0.0);
+    scale = tidx.x < 16 ? w_max_l : w_max_r;
+  }
+  scale /= bits == 4 ? 6.0f : 448.0f;
+
+  using ScaleType = metal::conditional_t<use_mx_scale, fp8_e8m0, fp8_e4m3>;
+  auto s = ScaleType(scale);
+  uint8_t q_scale = s.bits;
+  scale = float(s);
+
+  // Write out the scales and biases
+  size_t gindex = index / group_size;
+  if (index % group_size == 0) {
+    scales[gindex] = q_scale;
+  }
+
+  uint8_t output = Quantize<bits>{}(scale == 0 ? 0.0f : w_thread / scale);
+  if (bits == 4) {
+    uint8_t sval = simd_shuffle_down(output, 1);
+    output |= sval << bits;
+  }
+  constexpr int pack_factor = bits == 8 ? 1 : 2;
+  if (index % pack_factor == 0) {
+    out[index / pack_factor] = output;
+  }
+}
+
+template <typename T, const int group_size, const int bits>
+[[kernel]] void fp_dequantize(
+    const device uint8_t* w [[buffer(0)]],
+    const device T* scales [[buffer(1)]],
+    device T* out [[buffer(3)]],
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]) {
+  constexpr bool use_mx_scale = group_size == 32;
+  constexpr int pack_factor = bits == 8 ? 1 : 2;
+  size_t offset = index.x + grid_dim.x * size_t(index.y);
+  size_t oindex = offset * pack_factor;
+  size_t gindex = oindex / group_size;
+
+  out += oindex;
+
+  using ScaleType = metal::conditional_t<use_mx_scale, fp8_e8m0, fp8_e4m3>;
+  auto q_scale = ((device ScaleType*)(scales))[gindex];
+  auto scale = float(q_scale);
+
+  uint val = w[offset];
+#pragma clang loop unroll(full)
+  for (int i = 0; i < pack_factor; i++) {
+    uint8_t d;
+    if (bits == 4) {
+      d = (val >> (bits * i)) & 0x0f;
+    } else if (bits == 8) {
+      d = val;
+    }
+    out[i] = static_cast<T>(scale * Dequantize<bits>{}(d));
+  }
+}

mlx/backend/metal/kernels/fp_quantized.metal

@@ -4,7 +4,9 @@
 #include "mlx/backend/metal/kernels/utils.h"
 #include "mlx/backend/metal/kernels/steel/gemm/gemm.h"
 #include "mlx/backend/metal/kernels/quantized_utils.h"
-#include "mlx/backend/metal/kernels/fp4_quantized.h"
+#include "mlx/backend/metal/kernels/fp8.h"
+#include "mlx/backend/metal/kernels/fp4.h"
+#include "mlx/backend/metal/kernels/fp_quantized.h"
 
 #define instantiate_quantized(name, type) \
   instantiate_kernel( \
@@ -113,13 +115,33 @@
   instantiate_gather_qmm_rhs(mxfp4_gather_qmm_rhs, mxfp4_gather_qmm_rhs_nt, type, 16, 32, 32, 1, 2, true) \
   instantiate_gather_qmm_rhs(mxfp4_gather_qmm_rhs, mxfp4_gather_qmm_rhs_nn, type, 16, 32, 32, 1, 2, false)
 
+#define instantiate_quantize_dequantize(type, mode, group_size, bits) \
+  instantiate_kernel( \
+    mode "_quantize_" #type "_gs_" #group_size "_b_" #bits, \
+    fp_quantize, \
+    type, \
+    group_size,  \
+    bits) \
+  instantiate_kernel( \
+    mode "_dequantize_" #type "_gs_" #group_size "_b_" #bits, \
+    fp_dequantize, \
+    type, \
+    group_size,  \
+    bits)
+
+#define instantiate_quantize_dequantize_modes(type) \
+  instantiate_quantize_dequantize(type, "mxfp4", 32, 4) \
+  instantiate_quantize_dequantize(type, "nvfp4", 16, 4) \
+  instantiate_quantize_dequantize(type, "mxfp8", 32, 8)
+
 #define instantiate_quantized_types(type) \
   instantiate_quantized_all_batched(type) \
   instantiate_quantized_all_quad(type)    \
   instantiate_quantized_all_splitk(type)  \
   instantiate_quantized_all_single(type)  \
   instantiate_quantized_all_aligned(type) \
-  instantiate_quantized_all_rhs(type)
+  instantiate_quantized_all_rhs(type)     \
+  instantiate_quantize_dequantize_modes(type)
 
 instantiate_quantized_types(float)
 instantiate_quantized_types(bfloat16_t)

mlx/backend/metal/kernels/unary_ops.h

@@ -8,6 +8,7 @@
 #include "mlx/backend/metal/kernels/cexpf.h"
 #include "mlx/backend/metal/kernels/erf.h"
 #include "mlx/backend/metal/kernels/expm1f.h"
+#include "mlx/backend/metal/kernels/fp8.h"
 
 namespace {
 constant float inf = metal::numeric_limits<float>::infinity();
@@ -439,63 +440,15 @@ complex64_t ArcTan::operator()(complex64_t x) {
   return (1.0 / complex64_t{0.0, 2.0}) * Log{}((1.0 + ix) / (1.0 - ix));
 };
 
-inline float fp32_from_bits(uint32_t bits) {
-  return *(reinterpret_cast<thread float*>(&bits));
-}
-inline float fp32_to_bits(float x) {
-  return *(reinterpret_cast<thread uint32_t*>(&x));
-}
-
 struct ToFP8 {
   template <typename T>
   uint8_t operator()(T f) {
-    // From PyTorch
+    return fp8_e4m3(f).bits;
-    // https://github.com/pytorch/pytorch/blob/e3643e1e0e923f0fc063dfab6f45c956d568919d/c10/util/Float8_e4m3fn.h#L148
-    uint32_t fp8_max = 1087 << 20;
-    uint32_t denorm_mask = 141 << 23;
-    uint32_t f_bits = fp32_to_bits(static_cast<float>(f));
-    uint8_t result = 0u;
-    uint32_t sign = f_bits & 0x80000000;
-    f_bits ^= sign;
-    if (f_bits >= fp8_max) {
-      // Default behavior saturates to min/max
-      result = 0x7E;
-    } else {
-      if (f_bits < (121 << 23)) {
-        f_bits =
-            fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask));
-        result = static_cast<uint8_t>(f_bits - denorm_mask);
-      } else {
-        // resulting mantissa is odd
-        uint8_t mant_odd = (f_bits >> 20) & 1;
-        f_bits += ((uint32_t)(7 - 127) << 23) + 0x7FFFF;
-        f_bits += mant_odd;
-        result = static_cast<uint8_t>(f_bits >> 20);
-      }
-    }
-    result |= static_cast<uint8_t>(sign >> 24);
-    return result;
   }
 };
 
 struct FromFP8 {
   float operator()(uint8_t x) {
-    // From PyTorch:
+    return float(*(thread fp8_e4m3*)(&x));
-    // https://github.com/pytorch/pytorch/blob/e3643e1e0e923f0fc063dfab6f45c956d568919d/c10/util/Float8_e4m3fn.h#L46
-    uint32_t w = static_cast<uint32_t>(x) << 24;
-    uint32_t sign = w & 0x80000000;
-    uint32_t nonsign = w & 0x7FFFFFFF;
-
-    uint32_t renorm_shift = metal::clz(nonsign);
-    renorm_shift = renorm_shift > 4 ? renorm_shift - 4 : 0;
-
-    int32_t inf_nan_mask =
-        (static_cast<int32_t>(nonsign + 0x01000000) >> 8) & 0x7F800000;
-    int32_t zero_mask = static_cast<int32_t>(nonsign - 1) >> 31;
-    uint32_t result = sign |
-        ((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) |
-          inf_nan_mask) &
-         ~zero_mask);
-    return fp32_from_bits(result);
   }
 };

mlx/backend/metal/quantized.cpp

@@ -1045,26 +1045,31 @@ void fast::Quantize::eval_gpu(
   compute_encoder.set_input_array(w, 0);
   if (dequantize_) {
     auto scales = ensure_row_contiguous(inputs[1], d, s);
-    auto biases = ensure_row_contiguous(inputs[2], d, s);
     compute_encoder.set_input_array(scales, 1);
-    compute_encoder.set_input_array(biases, 2);
     compute_encoder.set_output_array(out, 3);
+    if (mode_ == QuantizationMode::Affine) {
+      auto biases = ensure_row_contiguous(inputs[2], d, s);
+      compute_encoder.set_input_array(biases, 2);
+    }
   } else {
     auto& scales = outputs[1];
-    auto& biases = outputs[2];
     scales.set_data(allocator::malloc(scales.nbytes()));
-    biases.set_data(allocator::malloc(biases.nbytes()));
     compute_encoder.set_output_array(out, 1);
     compute_encoder.set_output_array(scales, 2);
-    compute_encoder.set_output_array(biases, 3);
+    if (mode_ == QuantizationMode::Affine) {
+      auto& biases = outputs[2];
+      biases.set_data(allocator::malloc(biases.nbytes()));
+      compute_encoder.set_output_array(biases, 3);
+    }
   }
 
   auto type_string = dequantize_ ? get_type_string(out.dtype())
                                  : get_type_string(w_pre.dtype());
+  auto mode = quantization_mode_to_string(mode_);
   std::string kname;
   concatenate(
       kname,
-      dequantize_ ? "affine_dequantize" : "affine_quantize",
+      mode + (dequantize_ ? "_dequantize" : "_quantize"),
       "_",
       type_string,
       "_gs_",
@@ -1075,7 +1080,7 @@ void fast::Quantize::eval_gpu(
       d,
       kname,
       dequantize_ ? "dequantize" : "quantize",
-      "affine",
+      mode,
       type_string,
       group_size_,
       bits_);
@@ -1088,7 +1093,8 @@ void fast::Quantize::eval_gpu(
   int packs_per_int = (bits_ == 3 || bits_ == 5) ? 8
       : bits_ == 6                               ? 4
                                                  : 8 / bits_;
-  int per_thread = dequantize_ ? packs_per_int : group_size_ / simd_size;
+  int per_thread =
+      dequantize_ ? packs_per_int : std::max(group_size_ / simd_size, 1);
   size_t nthreads =
       dequantize_ ? out.size() / packs_per_int : w.size() / per_thread;
 

mlx/ops.cpp

@@ -2,7 +2,6 @@
 
 // Required for using M_PI in MSVC.
 #define _USE_MATH_DEFINES
-
 #include <algorithm>
 #include <climits>
 #include <cmath>
@@ -4259,8 +4258,11 @@ std::vector<array> fp_quantize(
     } else {
       // convert to e8m0
       auto z = array(0, scales.dtype());
-      scales =
+      scales = where(
-          where(equal(scales, z, s), z, astype(log2(scales, s), int32, s), s);
+          equal(scales, z, s),
+          z,
+          astype(round(log2(scales, s), s), int32, s),
+          s);
 
       wq = divide(wq, power(array(2.0f, w.dtype()), scales, s), s);
       scales = astype(add(scales, array(127, int32), s), uint8, s);

python/tests/test_quantized.py

@@ -92,7 +92,6 @@ class TestQuantized(mlx_tests.MLXTestCase):
 
         with self.assertRaises(ValueError):
             mx.quantize(w, group_size=32, bits=7, mode="mxfp8")
-
         w_q, scales = mx.quantize(w, group_size=32, bits=8, mode="mxfp8")
 
         with self.assertRaises(ValueError):
@@ -102,7 +101,8 @@ class TestQuantized(mlx_tests.MLXTestCase):
             mx.dequantize(w_q, scales, group_size=32, bits=4, mode="mxfp8")
 
         w_hat = mx.dequantize(w_q, scales, group_size=32, bits=8, mode="mxfp8")
-        self.assertTrue(mx.allclose(w, w_hat, rtol=1e-1, atol=1e-2))
+
+        self.assertTrue(mx.allclose(w, w_hat, rtol=1e-1, atol=1e-1))
 
         # test quantize/dequantize 0s
         a = mx.zeros((256, 512))