diff --git a/backends/vulkan/runtime/gen_vulkan_spv.py b/backends/vulkan/runtime/gen_vulkan_spv.py
index 5701248266e..76999cb0995 100644
--- a/backends/vulkan/runtime/gen_vulkan_spv.py
+++ b/backends/vulkan/runtime/gen_vulkan_spv.py
@@ -355,6 +355,38 @@ def layout_declare_buffer(
     return out_str
 
 
+# Vulkan leaves the rounding of a store to an rgba16f image to the device, and
+# some devices round toward zero. Rounding to the nearest f16, ties to even, in
+# the shader makes the store exact. Integer texels are stored as is, but stores
+# to rgba32f images would be rounded too, so a shader may not write both.
+ROUND_F16_IMAGE_STORES = """
+#ifdef WRITES_RGBA32F_IMAGE
+#error "A shader cannot write both rgba16f and rgba32f images"
+#endif
+#ifndef WRITES_RGBA16F_IMAGE
+#define WRITES_RGBA16F_IMAGE
+vec4 round_to_f16(const vec4 v) {
+  const uvec4 u = floatBitsToUint(v);
+  // Keep 10 mantissa bits: add just under half an f16 ulp, plus one when the
+  // kept lowest bit is odd so ties go to even, then clear the 13 dropped bits.
+  vec4 r = uintBitsToFloat((u + 0xFFFu + ((u >> 13) & 1u)) & ~0x1FFFu);
+  // Below the smallest normal f16, round to a multiple of 2^-24.
+  r = mix(r, roundEven(v * 16777216.0) * 5.9604644775390625e-8, lessThan(abs(v), vec4(6.103515625e-5)));
+  r = mix(r, uintBitsToFloat((u & 0x80000000u) | 0x7F800000u), greaterThan(abs(r), vec4(65504.0)));
+  return mix(r, v, isnan(v));
+}
+ivec4 round_to_f16(const ivec4 v) { return v; }
+uvec4 round_to_f16(const uvec4 v) { return v; }
+#define imageStore(image, pos, texel) imageStore(image, pos, round_to_f16(texel))
+#endif"""
+
+WRITES_RGBA32F_IMAGE = """
+#ifdef WRITES_RGBA16F_IMAGE
+#error "A shader cannot write both rgba16f and rgba32f images"
+#endif
+#define WRITES_RGBA32F_IMAGE"""
+
+
 def layout_declare_image(
     slot: Union[int, List[int]],
     access_type: str,
@@ -367,6 +399,11 @@ def layout_declare_image(
     image_type = TYPE_MAPPINGS["IMAGE_T"][image_ndim][dtype]
 
     ret_str = f"layout(set = 0, binding = {get_slot_val(slot)}, {image_format}) uniform {precision} restrict {get_access_qualifier(access_type)} {image_type} {var_name};"
+    if access_type.lower() != "r":
+        if image_format == "rgba16f":
+            ret_str += ROUND_F16_IMAGE_STORES
+        elif image_format == "rgba32f":
+            ret_str += WRITES_RGBA32F_IMAGE
 
     if isinstance(slot, list):
         slot[0] = slot[0] + 1
diff --git a/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_dw_prepack_weights.glsl b/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_dw_prepack_weights.glsl
index 4707bcf9f9c..4a16ee05d42 100644
--- a/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_dw_prepack_weights.glsl
+++ b/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_dw_prepack_weights.glsl
@@ -21,7 +21,7 @@ ${define_required_extensions("buffer", BUF_DTYPE)}
 
 layout(std430) buffer;
 
-layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[2][DTYPE]} image_out;
+${layout_declare_tensor(0, "w", "image_out", DTYPE, "texture2d")}
 layout(set = 0, binding = 1) buffer  PRECISION restrict readonly Buffer {
   BUF_T buffer_in[];
 };
diff --git a/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_prepack_weights.glsl b/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_prepack_weights.glsl
index b5bec8fca0b..95c8cc3c4ee 100644
--- a/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_prepack_weights.glsl
+++ b/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/pack_weights/conv2d_prepack_weights.glsl
@@ -29,7 +29,7 @@ $if OUTPUT_BUFFER:
     f16vec4 buffer_out[];
   };
 $else:
-  layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[2][DTYPE]} image_out;
+  ${layout_declare_tensor(0, "w", "image_out", DTYPE, "texture2d")}
 layout(set = 0, binding = 1) buffer  PRECISION restrict readonly Buffer {
   BUF_T buffer_in[];
 };
diff --git a/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/transposed/pack_weights/conv_transpose2d_prepack_weights.glsl b/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/transposed/pack_weights/conv_transpose2d_prepack_weights.glsl
index 65b998a9c5b..4a47892375e 100644
--- a/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/transposed/pack_weights/conv_transpose2d_prepack_weights.glsl
+++ b/backends/vulkan/runtime/graph/ops/glsl/convolution/conv2d/transposed/pack_weights/conv_transpose2d_prepack_weights.glsl
@@ -21,7 +21,7 @@ ${define_required_extensions("buffer", BUF_DTYPE)}
 
 layout(std430) buffer;
 
-layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[2][DTYPE]} image_out;
+${layout_declare_tensor(0, "w", "image_out", DTYPE, "texture2d")}
 layout(set = 0, binding = 1) buffer  PRECISION restrict readonly Buffer {
   BUF_T buffer_in[];
 };
diff --git a/backends/vulkan/runtime/graph/ops/glsl/repeat_channel.glsl b/backends/vulkan/runtime/graph/ops/glsl/repeat_channel.glsl
index 42c7f86aea8..96e9f7aac4c 100644
--- a/backends/vulkan/runtime/graph/ops/glsl/repeat_channel.glsl
+++ b/backends/vulkan/runtime/graph/ops/glsl/repeat_channel.glsl
@@ -16,7 +16,7 @@ layout(std430) buffer;
 
 #include "indexing_utils.h"
 
-layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[NDIM][DTYPE]} image_out;
+${layout_declare_image(0, "w", "image_out", DTYPE, image_ndim=NDIM)}
 layout(set = 0, binding = 1) uniform PRECISION sampler3D image_in;
 
 layout(set = 0, binding = 2) uniform PRECISION restrict RepeatArgs {
diff --git a/backends/vulkan/test/glsl/binary_op_nobroadcast__test.glsl b/backends/vulkan/test/glsl/binary_op_nobroadcast__test.glsl
index 7f72ac58972..5fcd1513c4c 100644
--- a/backends/vulkan/test/glsl/binary_op_nobroadcast__test.glsl
+++ b/backends/vulkan/test/glsl/binary_op_nobroadcast__test.glsl
@@ -14,9 +14,7 @@
 
 layout(std430) buffer;
 
-// clang-format off
-layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly image3D image_out;
-// clang-format on
+${layout_declare_image(0, "w", "image_out", DTYPE)}
 layout(set = 0, binding = 1) uniform PRECISION sampler3D image_in;
 layout(set = 0, binding = 2) uniform PRECISION sampler3D image_other;
 
diff --git a/backends/vulkan/test/glsl/fill_texture__test.glsl b/backends/vulkan/test/glsl/fill_texture__test.glsl
index 76c630de55e..637c12fbb51 100644
--- a/backends/vulkan/test/glsl/fill_texture__test.glsl
+++ b/backends/vulkan/test/glsl/fill_texture__test.glsl
@@ -12,7 +12,7 @@
 
 layout(std430) buffer;
 
-layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[NDIM][DTYPE]} uOutput;
+${layout_declare_image(0, "w", "uOutput", DTYPE, image_ndim=NDIM)}
 layout(set = 0, binding = 1) uniform PRECISION restrict Block {
   ivec3 size;
   int fill;
diff --git a/backends/vulkan/test/spv_codegen/test_gen_vulkan_spv.py b/backends/vulkan/test/spv_codegen/test_gen_vulkan_spv.py
index aef464dbf38..e569d6fe27d 100644
--- a/backends/vulkan/test/spv_codegen/test_gen_vulkan_spv.py
+++ b/backends/vulkan/test/spv_codegen/test_gen_vulkan_spv.py
@@ -13,6 +13,22 @@
 from runtime import gen_vulkan_spv
 
 
+class TestF16ImageStores(unittest.TestCase):
+    def test_writable_rgba16f_images_round_stores(self) -> None:
+        for access in ("w", "rw"):
+            with self.subTest(access=access):
+                declaration = gen_vulkan_spv.layout_declare_image(
+                    0, access, "t_out", "half"
+                )
+                self.assertIn("round_to_f16(texel)", declaration)
+
+    def test_other_images_store_unchanged(self) -> None:
+        for access, dtype in (("r", "half"), ("w", "float"), ("w", "int32")):
+            with self.subTest(access=access, dtype=dtype):
+                declaration = gen_vulkan_spv.layout_declare_image(0, access, "t", dtype)
+                self.assertNotIn("round_to_f16", declaration)
+
+
 class TestNestedShaderSources(unittest.TestCase):
     def _write_fake_compiler(
         self, directory: Path, assertions: str = ""
diff --git a/backends/vulkan/test/vulkan_compute_api_test.cpp b/backends/vulkan/test/vulkan_compute_api_test.cpp
index b087b0e4baa..7f4ef389040 100644
--- a/backends/vulkan/test/vulkan_compute_api_test.cpp
+++ b/backends/vulkan/test/vulkan_compute_api_test.cpp
@@ -1202,6 +1202,71 @@ TEST_F(VulkanComputeAPITest, spec_var_shader_test) {
   }
 }
 
+TEST_F(VulkanComputeAPITest, half_texture_store_rounds_to_nearest_even) {
+  // Ties, their neighbors, f16 subnormals and values at and past the largest
+  // f16, where rounding toward zero differs from rounding to nearest even.
+  const std::vector<float> values = {
+      1.0f + 0x1p-11f,
+      1.0f + 0x3p-11f,
+      -(1.0f + 0x3p-11f),
+      std::nextafter(1.0f + 0x1p-11f, 0.0f),
+      std::nextafter(1.0f + 0x1p-11f, 2.0f),
+      0x1p-25f,
+      0x3p-25f,
+      std::nextafter(0x1p-25f, 0.0f),
+      0x1.ffcp-15f,
+      -0x1.0p-24f,
+      65504.0f,
+      65519.0f,
+      65520.0f,
+      -65520.0f,
+      1e6f,
+      -1e6f,
+  };
+
+  struct Params final {
+    utils::ivec3 size;
+    int32_t fill;
+    utils::vec4 values;
+  };
+
+  for (size_t i = 0; i < values.size(); i += 4) {
+    vTensor a(
+        context(),
+        {4, 1, 1},
+        vkapi::kHalf,
+        utils::kTexture3D,
+        utils::kChannelsPacked);
+    Params block{
+        {1, 1, 1},
+        0,
+        {values[i], values[i + 1], values[i + 2], values[i + 3]},
+    };
+    ParamsBuffer params(context(), block);
+
+    vkapi::PipelineBarrier pipeline_barrier{};
+    context()->submit_compute_job(
+        VK_KERNEL(fill_texture__test_half),
+        pipeline_barrier,
+        GlobalWorkGrid({1u, 1u, 1u}, kExplicitWorkGrid),
+        LocalWorkGroup(1u, 1u, 1u),
+        {},
+        VK_NULL_HANDLE,
+        0,
+        a.image(
+            pipeline_barrier,
+            vkapi::PipelineStage::COMPUTE,
+            vkapi::MemoryAccessType::WRITE),
+        params.buffer());
+
+    const std::vector<float> out = extract_vtensor(a);
+    for (size_t j = 0; j < 4; ++j) {
+      EXPECT_EQ(out[j], float(executorch::aten::Half(values[i + j])))
+          << "input " << values[i + j];
+    }
+  }
+}
+
 TEST_F(VulkanComputeAPITest, update_params_between_submit) {
   context()->set_cmd(/*reusable = */ true);
   std::vector<int64_t> sizes = {4, 4, 2};