[ET-VK] VMA assertion crash on macOS/MoltenVK: test_host_cached_available() returns bool instead of VmaAllocationCreateFlags

[abdelaziz-mahdy]

Bug

test_host_cached_available() in backends/vulkan/runtime/vk_api/memory/Allocator.cpp is declared as returning bool but returns VmaAllocationCreateFlags values, causing implicit truncation to 1 (true).

This results in allocation_strategy_device_to_host_ being set to 1 (VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT) instead of the intended VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT (0x800) or VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT (0x400).

When a DEVICE_TO_HOST staging buffer is created, the invalid flag causes a VMA assertion failure:

Assertion failed: ((inoutCreateInfo.flags & (VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT | 
VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT)) != 0 && "When using VMA_ALLOCATION_CREATE_MAPPED_BIT 
and usage = VMA_MEMORY_USAGE_AUTO*, you must also specify VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT 
or VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT."), function CalcAllocationParams, file vk_mem_alloc.h, line 14926.

Reproduction

• Platform: macOS (Apple Silicon) with MoltenVK
• Model: Any Vulkan model exported with force_fp16=True
• ExecuTorch version: 1.2.0

Works fine on 1.1.0 where this function didn't exist.

Root Cause

// Allocator.cpp, line 13
bool test_host_cached_available(VkPhysicalDevice physical_device) {  // <- returns bool
  // ...
  if (host_visible && host_cached) {
    return VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;       // 0x800 -> truncated to true (1)
  }
  return VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT; // 0x400 -> truncated to true (1)
}

Then at line 37-38:

allocation_strategy_device_to_host_{test_host_cached_available(physical_device_)}
// Gets value 1 instead of 0x800 or 0x400

Used at line 181:

alloc_create_info.flags |= allocation_strategy_device_to_host_;  // ORs in 1 instead of a host access flag

Fix

Change the return type from bool to VmaAllocationCreateFlags:

VmaAllocationCreateFlags test_host_cached_available(VkPhysicalDevice physical_device) {

[abdelaziz-mahdy]

This is already fixed on main — the function now returns VmaAllocationCreateFlags and properly iterates all memory types. The fix didn't make it into the v1.2.0 release tag. Could this be cherry-picked into a 1.2.1 patch release?

since it crashes when loading vulkan models.

[abdelaziz-mahdy]

Bug introduced by: PR #17856 ("Fix for Raspberry Pi 5 GPU") — cherry-picked into v1.2.0 release branch with bool return type on test_host_cached_available().

Fixed on main by:

• PR [ET-VK] Fix VMA assertion caused by bool return type truncation #18105 ("[ET-VK] Fix VMA assertion caused by bool return type truncation") — merged Mar 11, 2 days after the bug
• PR [ET-VK] Fix staging buffer allocation to check all memory types for HOST_CACHED #18291 ("[ET-VK] Fix staging buffer allocation to check all memory types for HOST_CACHED") — further improved to iterate all memory types

What happened: PR #17856 was cherry-picked into the v1.2.0 release branch, but the follow-up fixes (#18105, #18291) were not. The v1.2.0 tag ships with the buggy bool return type.

Impact: Any Vulkan model loading on macOS/MoltenVK crashes with VMA assertion failure. Also affects any device where DEVICE_TO_HOST staging buffers are created.

Could #18105 and #18291 be cherry-picked into a v1.2.1 patch release?

@SS-JIA what do you think?