UPSTREAM PR #18302: vulkan: fix command buffer corruption in ggml_backend_vk_event_wait

[loci-dev]

Mirrored from ggml-org/llama.cpp#18302

Fixes #18297.

The async download paths that I had been using to test events weren't hitting event_wait, and this bug slipped through.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #667

PR Context: Vulkan backend bug fix for command buffer corruption in ggml_backend_vk_event_wait

Analysis Result

This PR introduces a correctness fix in the Vulkan backend event synchronization path. The changes add proper cleanup of transfer context resources and debug logging to three functions in ggml-vulkan.cpp:

• ggml_backend_vk_event_record
• ggml_backend_vk_event_wait
• ggml_backend_vk_device_event_synchronize

Code Changes:

• Added ggml_vk_ctx_end(transfer_ctx) call to properly end the transfer context
• Added ctx->transfer_ctx.reset() to release the context resources
• Added debug logging statements for event tracking

Performance Impact:
No measurable performance changes detected. The power consumption analysis shows 0% change across all stable binaries. The core inference functions (llama_decode, llama_encode, llama_tokenize) are not modified. Token throughput remains unaffected as this fix operates in the Vulkan backend event synchronization layer, which is outside the primary inference path. The added cleanup operations execute only during event wait operations and introduce negligible overhead compared to GPU command submission latency.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #667

PR Context: Vulkan backend bug fix for command buffer corruption in ggml_backend_vk_event_wait

Analysis Result

This PR introduces a correctness fix in the Vulkan backend event synchronization path. The changes add proper cleanup of transfer context resources and debug logging to three functions in ggml-vulkan.cpp:

• ggml_backend_vk_event_record
• ggml_backend_vk_event_wait
• ggml_backend_vk_device_event_synchronize

Code Changes:

• Added ggml_vk_ctx_end(transfer_ctx) call to properly end the transfer context
• Added ctx->transfer_ctx.reset() to release the context resources
• Added debug logging statements for event tracking

Performance Impact:
No measurable performance changes detected. The power consumption analysis shows 0% change across all stable binaries. The core inference functions (llama_decode, llama_encode, llama_tokenize) are not modified. Token throughput remains unaffected as this fix operates in the Vulkan backend event synchronization layer, which is outside the primary inference path. The added cleanup operations execute only during event wait operations and introduce negligible overhead compared to GPU command submission latency.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #667

PR Context: Vulkan backend bug fix for command buffer corruption in ggml_backend_vk_event_wait

Analysis Result

This PR introduces a correctness fix in the Vulkan backend event synchronization path. The changes add proper cleanup of transfer context resources and debug logging to three functions in ggml-vulkan.cpp:

• ggml_backend_vk_event_record
• ggml_backend_vk_event_wait
• ggml_backend_vk_device_event_synchronize

Code Changes:

• Added ggml_vk_ctx_end(transfer_ctx) call to properly end the transfer context
• Added ctx->transfer_ctx.reset() to release the context resources
• Added debug logging statements for event tracking

Performance Impact:
No measurable performance changes detected. The power consumption analysis shows 0% change across all stable binaries. The core inference functions (llama_decode, llama_encode, llama_tokenize) are not modified. Token throughput remains unaffected as this fix operates in the Vulkan backend event synchronization layer, which is outside the primary inference path. The added cleanup operations execute only during event wait operations and introduce negligible overhead compared to GPU command submission latency.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #667

PR Context: Vulkan backend bug fix for command buffer corruption in ggml_backend_vk_event_wait

Analysis Result

This PR introduces a correctness fix in the Vulkan backend event synchronization path. The changes add proper cleanup of transfer context resources and debug logging to three functions in ggml-vulkan.cpp:

• ggml_backend_vk_event_record
• ggml_backend_vk_event_wait
• ggml_backend_vk_device_event_synchronize

Code Changes:

• Added ggml_vk_ctx_end(transfer_ctx) call to properly end the transfer context
• Added ctx->transfer_ctx.reset() to release the context resources
• Added debug logging statements for event tracking

Performance Impact:
No measurable performance changes detected. The power consumption analysis shows 0% change across all stable binaries. The core inference functions (llama_decode, llama_encode, llama_tokenize) are not modified. Token throughput remains unaffected as this fix operates in the Vulkan backend event synchronization layer, which is outside the primary inference path. The added cleanup operations execute only during event wait operations and introduce negligible overhead compared to GPU command submission latency.