Update steel::AttnParams

Author: jagrit06

mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.h

@@ -1,4 +1,4 @@
-// Copyright © 2024 Apple Inc.
+// Copyright © 2024-25 Apple Inc.
 
 using namespace mlx::steel;
 
@@ -9,6 +9,9 @@ using namespace mlx::steel;
 constant bool align_Q [[function_constant(200)]];
 constant bool align_K [[function_constant(201)]];
 
+constant bool has_mask [[function_constant(300)]];
+constant bool do_causal [[function_constant(301)]];
+
 template <typename T>
 struct TransformScale {
   T scale;
@@ -69,13 +72,16 @@ template <
     int BD,
     int WM,
     int WN,
+    typename MaskType = float,
     typename AccumType = float>
 [[kernel, max_total_threads_per_threadgroup(WM * WN * 32)]] void attention(
     const device T* Q [[buffer(0)]],
     const device T* K [[buffer(1)]],
     const device T* V [[buffer(2)]],
     device T* O [[buffer(3)]],
     const constant AttnParams* params [[buffer(4)]],
+    const constant AttnMaskParams* mask_params [[buffer(5), function_constant(has_mask)]],
+    const device MaskType* mask [[buffer(6), function_constant(has_mask)]],
     uint simd_lane_id [[thread_index_in_simdgroup]],
     uint simd_group_id [[simdgroup_index_in_threadgroup]],
     uint3 tid [[threadgroup_position_in_grid]],
@@ -102,6 +108,12 @@ template <
       tidl.y * params->O_strides[1] + // Head
       tidl.x * BQ * params->O_strides[2]; // Seqeunce
 
+  if (has_mask) {
+    mask += tidl.z * mask_params->M_strides[0] + // Batch
+        tidl.y * mask_params->M_strides[1] + // Head
+        tidl.x * BQ * mask_params->M_strides[2]; // Seqeunce
+  }
+
   // Prepare threadgroup memory
   constexpr short padQ = 16 / sizeof(T);
   constexpr short padK = 16 / sizeof(T);
@@ -203,7 +215,7 @@ template <
 
   // Load Q blocks apply scale
   if (!align_Q && int(tid.x) == (params->NQ_aligned)) {
-    loader_q.load_safe(short2(BD, params->qL - params->NQ_aligned * BQ));
+    loader_q.load_safe(short2(BD, params->qL_rem));
   } else {
     loader_q.load_unsafe();
   }
@@ -226,7 +238,7 @@ template <
     // Load K block and apply scale
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (!align_K && kb == (params->NK_aligned)) {
-      loader_k.load_safe(short2(BD, params->kL - params->NK_aligned * BK));
+      loader_k.load_safe(short2(BD, params->kL_rem));
     } else {
       loader_k.load_unsafe();
     }
@@ -276,7 +288,7 @@ template <
 
     // Load V blocks
     if (!align_K && kb == (params->NK_aligned)) {
-      loader_v.load_safe(short2(BD, params->kL - params->NK_aligned * BK));
+      loader_v.load_safe(short2(BD, params->kL_rem));
     } else {
       loader_v.load_unsafe();
     }
@@ -367,8 +379,7 @@ template <
   O += (tm + sm) * params->O_strides[2] + sn;
 
   if (!align_Q && int(tid.x) == (params->NQ_aligned)) {
-    auto dst_tile_dims =
-        short2(BD - sn, params->qL - BQ * params->NQ_aligned - (tm + sm));
+    auto dst_tile_dims = short2(BD - sn, params->qL_rem - (tm + sm));
 
     if (dst_tile_dims.x <= 0 || dst_tile_dims.y <= 0)
       return;

mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.metal

@@ -1,31 +1,28 @@
-// Copyright © 2024 Apple Inc.
+// Copyright © 2024-25 Apple Inc.
 
 // clang-format off
 #include "mlx/backend/metal/kernels/utils.h"
 
 #include "mlx/backend/metal/kernels/steel/attn/attn.h"
 #include "mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.h"
 
-#define instantiate_attn(tname, dtype, bq, bk, bd, wm, wn) \
-  template [[host_name("steel_attention_" #tname "_bq" #bq "_bk" #bk "_bd" #bd "_wm" #wm "_wn" #wn)]] \
-  [[kernel]] void attention<dtype, bq, bk, bd, wm, wn, float>( \
-      const device dtype* Q [[buffer(0)]], \
-      const device dtype* K [[buffer(1)]], \
-      const device dtype* V [[buffer(2)]], \
-      device dtype* O [[buffer(3)]],\
-      const constant AttnParams* params [[buffer(4)]], \
-      uint simd_lane_id [[thread_index_in_simdgroup]], \
-      uint simd_group_id [[simdgroup_index_in_threadgroup]], \
-      uint3 tid [[threadgroup_position_in_grid]], \
-      uint3 lid [[thread_position_in_threadgroup]]);
+#define instantiate_attn(tname, dtype, bq, bk, bd, wm, wn, mname, mtype) \
+  instantiate_kernel(                                                    \
+      "steel_attention_" #tname "_bq" #bq "_bk" #bk "_bd" #bd            \
+      "_wm" #wm "_wn" #wn "_mask" #mname,                                \
+  attention, dtype, bq, bk, bd, wm, wn, mtype, float)
 
-#define instantiate_attn_shapes_helper(iname, itype) \
-    instantiate_attn(iname, itype, 32, 16, 128, 4, 1) \
-    instantiate_attn(iname, itype, 32, 32,  80, 4, 1) \
-    instantiate_attn(iname, itype, 32, 32,  64, 4, 1)
+#define instantiate_attn_shapes_helper(iname, itype, mname, mtype)  \
+    instantiate_attn(iname, itype, 32, 16, 128, 4, 1, mname, mtype) \
+    instantiate_attn(iname, itype, 32, 32,  80, 4, 1, mname, mtype) \
+    instantiate_attn(iname, itype, 32, 32,  64, 4, 1, mname, mtype)
 
-instantiate_attn_shapes_helper(float16, half);
-instantiate_attn_shapes_helper(bfloat16, bfloat16_t);
+#define instantiate_attn_mask_helper(iname, itype) \
+    instantiate_attn_shapes_helper(iname, itype, iname, itype) \
+    instantiate_attn_shapes_helper(iname, itype, bool_, bool)
 
-instantiate_attn_shapes_helper(float32, float);
+instantiate_attn_mask_helper(float16, half);
+instantiate_attn_mask_helper(bfloat16, bfloat16_t);
+
+instantiate_attn_mask_helper(float32, float);
 // clang-format on

mlx/backend/metal/kernels/steel/attn/params.h

@@ -26,11 +26,19 @@ struct AttnParams {
   int NQ_aligned; ///< Number of full query blocks
   int NK_aligned; ///< Number of full key/value blocks
 
+  int qL_rem; ///< Remainder in last query block
+  int kL_rem; ///< Remainder in last key/value block
+  int qL_off; ///< Offset in query sequence start
+
   int64_t Q_strides[3]; ///< Query  strides (B, H, L, D = 1)
   int64_t K_strides[3]; ///< Key    strides (B, H, L, D = 1)
   int64_t V_strides[3]; ///< Value  strides (B, H, L, D = 1)
   int64_t O_strides[3]; ///< Output strides (B, H, L, D = 1)
 };
 
+struct AttnMaskParams {
+  int64_t M_strides[3]; ///< Mask  strides (B, H, qL, kL = 1)
+};
+
 } // namespace steel
 } // namespace mlx

mlx/backend/metal/scaled_dot_product_attention.cpp

@@ -21,7 +21,9 @@ void sdpa_full_self_attention_metal(
     const array& k,
     const array& v,
     const float scale,
-    array& o) {
+    array& o,
+    bool do_causal_ = false,
+    const std::optional<array>& mask = std::nullopt) {
   using namespace mlx::steel;
 
   int wm = 4;
@@ -41,11 +43,14 @@ void sdpa_full_self_attention_metal(
 
   const bool align_Q = (qL % bq) == 0;
   const bool align_K = (kL % bk) == 0;
+  const bool has_mask = !!mask;
+  const bool do_causal = do_causal_;
 
   metal::MTLFCList func_consts = {
       {&align_Q, MTL::DataType::DataTypeBool, 200},
       {&align_K, MTL::DataType::DataTypeBool, 201},
-  };
+      {&has_mask, MTL::DataType::DataTypeBool, 300},
+      {&do_causal, MTL::DataType::DataTypeBool, 301}};
 
   std::ostringstream kname;
   // clang-format off
@@ -54,13 +59,17 @@ void sdpa_full_self_attention_metal(
         << "_bq" << bq
         << "_bk" << bk
         << "_bd" << bd
-        << "_wm" << wm << "_wn" << wn; // clang-format on
+        << "_wm" << wm 
+        << "_wn" << wn
+        << "_mask" << (type_to_name(mask ? *mask : q)); // clang-format on
 
   std::string base_name = kname.str();
 
   // clang-format off
   kname << "_align_Q_" << (align_Q ? 't' : 'n')
-        << "_align_K_" << (align_K ? 't' : 'n'); // clang-format on
+        << "_align_K_" << (align_K ? 't' : 'n') 
+        << "_has_mask_" << (has_mask ? 't' : 'n')
+        << "_do_causal_" << (do_causal ? 't' : 'n'); // clang-format on
 
   std::string hash_name = kname.str();
 
@@ -91,6 +100,10 @@ void sdpa_full_self_attention_metal(
       /* int NQ_aligned = */ NQ_aligned,
       /* int NK_aligned = */ NK_aligned,
 
+      /* int qL_rem = */ (qL - NQ_aligned * bq),
+      /* int kL_rem = */ (kL - NK_aligned * bk),
+      /* int qL_off = */ ((kL - qL) < 0 ? 0 : (kL - qL)),
+
       /* int64_t Q_strides[3] = */ {q.strides(0), q.strides(1), q.strides(2)},
       /* int64_t K_strides[3] = */ {k.strides(0), k.strides(1), k.strides(2)},
       /* int64_t V_strides[3] = */ {v.strides(0), v.strides(1), v.strides(2)},
@@ -102,6 +115,15 @@ void sdpa_full_self_attention_metal(
   compute_encoder.set_output_array(o, 3);
   compute_encoder.set_bytes(params, 4);
 
+  if (mask) {
+    auto m = *mask;
+    AttnMaskParams mask_params{/* int64_t M_strides[3] = */ {
+        m.strides(0), m.strides(1), m.strides(2)}};
+
+    compute_encoder.set_bytes(mask_params, 5);
+    compute_encoder.set_input_array(m, 6);
+  }
+
   MTL::Size grid_dims = MTL::Size(NQ, H, B);
   MTL::Size group_dims = MTL::Size(32, wm, wn);
 
@@ -324,7 +346,7 @@ void ScaledDotProductAttention::eval_gpu(
 
   // Checks that the headdim dimension has stride 1.
   auto is_matrix_contiguous = [](const array& arr) {
-    return arr.strides(3) == 1;
+    return arr.strides(-1) == 1;
   };
 
   // We are in vector mode ie single query
@@ -381,7 +403,11 @@ void ScaledDotProductAttention::eval_gpu(
         {str_oB, str_oH, str_oL, str_oD},
         flags);
 
-    sdpa_full_self_attention_metal(s, d, q, k, v, scale_, o);
+    auto mask = inputs.size() > 3
+        ? std::optional<array>{copy_unless(is_matrix_contiguous, inputs[3])}
+        : std::nullopt;
+
+    sdpa_full_self_attention_metal(s, d, q, k, v, scale_, o, do_causal_, mask);
   }
 
   d.add_temporaries(std::move(copies), s.index);

mlx/fast.cpp

@@ -710,7 +710,7 @@ array scaled_dot_product_attention(
         auto k_idx = arange(0, kL, s);
         q_idx = expand_dims(q_idx, 1, s);
         k_idx = expand_dims(k_idx, 0, s);
-        mask = less(q_idx, k_idx, s);
+        mask = greater_equal(q_idx, k_idx, s);
       }
 
       if (n_repeats > 1 && mask.ndim() >= 3) {
@@ -746,11 +746,15 @@ array scaled_dot_product_attention(
   const bool sdpa_full_supported_head_dim = query_head_dim == value_head_dim &&
       (query_head_dim == 64 || query_head_dim == 80 || query_head_dim == 128);
 
+  const bool sdpa_vector_supported_mask = (!has_mask || has_bool_mask);
+  const bool sdpa_full_supported_mask = !has_mask;
+
   const bool supports_sdpa_full = query_sequence_length >= threshold &&
-      !has_mask && sdpa_full_supported_head_dim && stream.device == Device::gpu;
+      sdpa_full_supported_mask && sdpa_full_supported_head_dim &&
+      stream.device == Device::gpu;
 
-  const bool supports_sdpa_vector = query_sequence_length == 1 &&
-      (!has_mask || has_bool_mask) && sdpa_vector_supported_head_dim &&
+  const bool supports_sdpa_vector = query_sequence_length < 8 &&
+      sdpa_vector_supported_mask && sdpa_vector_supported_head_dim &&
       stream.device == Device::gpu;
 
   const bool implementation_supports_use_case =

python/tests/test_fast_sdpa.py

@@ -30,7 +30,7 @@ def mlx_ref_attn(q, k, v, scale=1.0, mask=None):
             q_offset = max(0, kL - L)
             q_indices = mx.arange(q_offset, q_offset + L)
             k_indices = mx.arange(kL)
-            mask = q_indices[:, None] < k_indices[None]
+            mask = q_indices[:, None] >= k_indices[None]
 
         if n_repeats > 1 and mask.ndim >= 3:
             if mask.shape[-3] == 1: