Add a test

mlx/backend/metal/kernels/fp_quantized.h

@@ -533,6 +533,7 @@ METAL_FUNC void fp_qvm_impl(
     device T* y,
     const int in_vec_size,
     const int out_vec_size,
+    const int in_vec_stride,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
@@ -563,7 +564,7 @@ METAL_FUNC void fp_qvm_impl(
   int out_col = pack_factor * tn * (tid.y * num_simdgroups + simd_gid);
   ws += out_col * bytes_per_pack / pack_factor + simd_lid * out_vec_size_w;
   scales += out_col / group_size + simd_lid * out_vec_size_g;
-  x += tid.x * in_vec_size + simd_lid;
+  x += tid.x * in_vec_stride + simd_lid;
   y += tid.x * out_vec_size + out_col;
 
   if (out_col >= out_vec_size) {
@@ -1122,7 +1123,16 @@ template <typename T, const int group_size, int bits, bool batched>
         tid);
   }
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
+      w,
+      scales,
+      x,
+      y,
+      in_vec_size,
+      out_vec_size,
+      in_vec_size,
+      tid,
+      simd_gid,
+      simd_lid);
 }
 
 template <typename T, const int group_size, int bits, int split_k = 32>
@@ -1164,8 +1174,20 @@ template <typename T, const int group_size, int bits, int split_k = 32>
   int in_vec_size_adj =
       tid.z % split_k == split_k - 1 ? final_block_size : in_vec_size;
 
+  // The in_vec_stride is the full K dimension, not the partition size
+  int in_vec_stride = (split_k - 1) * in_vec_size + final_block_size;
+
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size_adj, out_vec_size, tid, simd_gid, simd_lid);
+      w,
+      scales,
+      x,
+      y,
+      in_vec_size_adj,
+      out_vec_size,
+      in_vec_stride,
+      tid,
+      simd_gid,
+      simd_lid);
 }
 
 template <
@@ -1423,7 +1445,16 @@ template <typename T, int group_size, int bits>
       s_strides,
       tid);
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
+      w,
+      scales,
+      x,
+      y,
+      in_vec_size,
+      out_vec_size,
+      in_vec_size,
+      tid,
+      simd_gid,
+      simd_lid);
 }
 
 template <

mlx/backend/metal/kernels/quantized.h

@@ -983,6 +983,7 @@ METAL_FUNC void qvm_impl(
     device T* y,
     const int in_vec_size,
     const int out_vec_size,
+    const int in_vec_stride,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
@@ -1016,7 +1017,7 @@ METAL_FUNC void qvm_impl(
   ws += out_col * bytes_per_pack / pack_factor + simd_lid * out_vec_size_w;
   scales += out_col / group_size + simd_lid * out_vec_size_g;
   biases += out_col / group_size + simd_lid * out_vec_size_g;
-  x += tid.x * in_vec_size + simd_lid;
+  x += tid.x * in_vec_stride + simd_lid;
   y += tid.x * out_vec_size + out_col;
 
   if (out_col >= out_vec_size) {
@@ -1643,6 +1644,7 @@ template <typename T, const int group_size, const int bits, bool batched>
       y,
       in_vec_size,
       out_vec_size,
+      in_vec_size,
       tid,
       simd_gid,
       simd_lid);
@@ -1691,6 +1693,9 @@ template <typename T, const int group_size, const int bits, int split_k = 32>
   int in_vec_size_adj =
       tid.z % split_k == split_k - 1 ? final_block_size : in_vec_size;
 
+  // The in_vec_stride is the full K dimension, not the partition size
+  int in_vec_stride = (split_k - 1) * in_vec_size + final_block_size;
+
   qvm_impl<T, group_size, bits>(
       w,
       scales,
@@ -1699,6 +1704,7 @@ template <typename T, const int group_size, const int bits, int split_k = 32>
       y,
       in_vec_size_adj,
       out_vec_size,
+      in_vec_stride,
       tid,
       simd_gid,
       simd_lid);
@@ -2077,6 +2083,7 @@ template <typename T, int group_size, int bits>
       y,
       in_vec_size,
       out_vec_size,
+      in_vec_size,
       tid,
       simd_gid,
       simd_lid);

mlx/backend/metal/rope.cpp

@@ -133,9 +133,8 @@ void RoPE::eval_gpu(
   if (single) {
     compute_encoder.set_bytes(out_strides, 1, 4);
     uint32_t dim0 = dims_ / 2;
-    uint32_t dim1 = B * N;
-    group_dims = get_block_dims(dim0, dim1, 1);
-    grid_dims = MTL::Size(dim0, dim1, 1);
+    group_dims = get_block_dims(dim0, N, 1);
+    grid_dims = MTL::Size(dim0, N, 1);
   } else {
     compute_encoder.set_bytes(strides, 3, 4);
     compute_encoder.set_bytes(out_strides, 3, 5);

python/tests/test_fast.py

@@ -365,28 +365,6 @@ class TestFast(mlx_tests.MLXTestCase):
         rx = rope_orig(x, dims, traditional, base, scale, offset)
         self.assertLess(mx.abs(rx - rx_fast).max(), 1e-5)
 
-    def test_rope_single_batch(self):
-        base = 10000.0
-        scale = 1.0
-        offset = 5
-
-        for traditional in [True, False]:
-            for B in [2, 4, 8]:
-                for n_head in [1, 4, 7]:
-                    for dims in [64, 128]:
-                        x = mx.random.uniform(shape=(B, n_head, 1, dims))
-                        mx.eval(x)
-                        rx_fast = mx.fast.rope(
-                            x,
-                            dims,
-                            traditional=traditional,
-                            base=base,
-                            scale=scale,
-                            offset=offset,
-                        )
-                        rx = rope_orig(x, dims, traditional, base, scale, offset)
-                        self.assertLess(mx.abs(rx - rx_fast).max(), 1e-5)
-
     def test_rope_with_large_offset(self):
         x = mx.random.normal(shape=(1, 1, 1024, 32))
         rx_fp32 = mx.fast.rope(

python/tests/test_quantized.py

@@ -470,6 +470,30 @@ class TestQuantized(mlx_tests.MLXTestCase):
         self.assertEqual(y_q.shape, y_hat.shape)
         self.assertLess((y_q - y_hat).abs().max(), 2e-3)
 
+    def test_qvm_splitk_multi_row(self):
+        # Test qvm split_k with M > 1 to ensure the x row stride is correct
+        key = mx.random.key(0)
+        k1, k2 = mx.random.split(key)
+        tests = product(
+            [64, 32],  # group_size
+            [4, 8],  # bits
+            [128],  # out dim (N)
+            [2048, 4096],  # in dim (K) >= 1024 to trigger split_k
+            [2, 3],  # M (multiple rows)
+        )
+        for group_size, bits, N, K, M in tests:
+            with self.subTest(M=M, K=K, N=N, group_size=group_size, bits=bits):
+                x = 1e-1 * mx.random.normal(shape=(M, K), key=k1)
+                w = 1e-1 * mx.random.normal(shape=(K, N), key=k2)
+                w_q, scales, biases = mx.quantize(w, group_size, bits)
+                w_hat = mx.dequantize(w_q, scales, biases, group_size, bits)
+                y_q = mx.quantized_matmul(
+                    x, w_q, scales, biases, False, group_size, bits
+                )
+                y_hat = x @ w_hat
+                self.assertEqual(y_q.shape, y_hat.shape)
+                self.assertLess((y_q - y_hat).abs().max(), 2e-3)
+
     def test_fp_qvm(self):
         key = mx.random.key(0)
         k1, k2 = mx.random.split(key)