Device API V2 Improvements #11150
Merged
yosefe merged 18 commits intoopenucx:masterfrom Mar 4, 2026
Merged
Conversation
| /* Migrate the constructed handle header to device memory */ | ||
| ucp_mem_type_unpack(ep->worker, mem->address, &handle, sizeof(handle), | ||
| ucp_mem_type_unpack(ep->worker, mem->address, handle, handle_size, | ||
| mem_type); |
| uct_cuda_ipc_device_mem_element_t cuda_ipc_mem_element; | ||
| }; | ||
|
|
||
| /* Base structure for all device memory elements */ |
| @@ -772,16 +767,13 @@ static ucs_status_t ucp_device_remote_mem_list_element_pack( | |||
| uct_rkey = ucp_rkey_get_tl_rkey(rkey, rkey_index); | |||
| ucs_assert(uct_rkey != UCT_INVALID_RKEY); | |||
|
|
|||
tvegas1
reviewed
Jan 29, 2026
tvegas1
previously approved these changes
Feb 4, 2026
tvegas1
reviewed
Feb 10, 2026
tvegas1
previously approved these changes
Feb 10, 2026
yosefe
reviewed
Feb 14, 2026
| UCS_F_DEVICE ucs_status_t ucp_device_check_params(const T mem_list_h, | ||
| unsigned index) | ||
| { | ||
| if (DEVICE_ENABLE_PARAMS_CHECK) { |
Contributor
There was a problem hiding this comment.
this macro should have UCP_ prefix since it's compiled as part of user's code
Comment on lines
+57
to
+58
| if (((mem_list_h)->version != UCP_DEVICE_MEM_LIST_VERSION_V1) || | ||
| ((index) >= (mem_list_h)->length)) { |
Contributor
There was a problem hiding this comment.
extra () not needed it's not a macro
| if (DEVICE_ENABLE_PARAMS_CHECK) { | ||
| if (((mem_list_h)->version != UCP_DEVICE_MEM_LIST_VERSION_V1) || | ||
| ((index) >= (mem_list_h)->length)) { | ||
| printf("Invalid parameters for %p\n", mem_list_h); |
| #define UCS_F_DEVICE static inline | ||
| #endif /* __NVCC__ */ | ||
|
|
||
| #ifndef DEVICE_ENABLE_PARAMS_CHECK |
| uint32_t lkey; | ||
| uint32_t rkey; | ||
| } uct_rc_gdaki_device_mem_element_t; | ||
| } uct_ib_md_device_mem_element_t; |
Contributor
There was a problem hiding this comment.
shall we move these structs to ib/cuda_ipc and not have them in uct common code?
Contributor
Author
There was a problem hiding this comment.
Will move in next PR after revert to query size from transport.
| } | ||
|
|
||
| static ucs_status_t | ||
| uct_ib_md_mlx5_devx_mem_elem_pack(uct_md_h md, uct_mem_h memh, uct_rkey_t rkey, |
Contributor
There was a problem hiding this comment.
uct_ib_md_mlx5_devx_md_mem_elem_pack
| { | ||
| size_t handle_size = 0; | ||
| size_t uct_elem_size[UCP_DEVICE_MEM_LIST_MAX_EPS]; | ||
| const size_t uct_elem_size = sizeof(uct_device_mem_element_t); |
Contributor
Author
There was a problem hiding this comment.
Do we want to touch code we plan to remove?
| return UCS_ERR_NO_DEVICE; | ||
| } | ||
|
|
||
| handle_size *= params->num_elements; |
Contributor
There was a problem hiding this comment.
maybe add variable like num_lanes, that we increment in line 381, and then
handle_size = sizeof(*handle) + (num_lanes * params->num_elements)
it's better that use same variable with different meanings over time
Contributor
Author
There was a problem hiding this comment.
Do we want to touch code we plan to remove?
Comment on lines
+490
to
+494
| handle->local_addrs = (void**)UCS_PTR_BYTE_OFFSET( | ||
| mem->address, UCS_PTR_BYTE_DIFF(handle, local_addresses)); | ||
| handle->remote_addrs = (uint64_t*)UCS_PTR_BYTE_OFFSET( | ||
| mem->address, UCS_PTR_BYTE_DIFF(handle, remote_addresses)); | ||
| handle->lengths = (size_t*)UCS_PTR_BYTE_OFFSET(mem->address, |
Contributor
There was a problem hiding this comment.
i guess we could remove these after removing v1 device api from ucx?
Comment on lines
+868
to
+880
| status = ucp_mem_do_alloc(context, NULL, handle_size, | ||
| UCT_MD_MEM_ACCESS_LOCAL_READ | | ||
| UCT_MD_MEM_ACCESS_LOCAL_WRITE, | ||
| mem_type, local_sys_dev, | ||
| "ucp_device_remote_mem_list_handle_t", mem); | ||
| if (status != UCS_OK) { | ||
| ucs_error("failed to allocate ucp_device_remote_mem_list_handle_t: %s", | ||
| ucs_status_string(status)); | ||
| goto out; | ||
| } | ||
|
|
||
| return UCS_OK; | ||
| ucp_mem_type_unpack(ep->worker, mem->address, handle, handle_size, | ||
| mem_type); |
Contributor
There was a problem hiding this comment.
can unite common code with create local mem list handle?
tvegas1
reviewed
Feb 17, 2026
tvegas1
previously approved these changes
Feb 25, 2026
yosefe
reviewed
Mar 1, 2026
yosefe
approved these changes
Mar 3, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What?
Fix comments and improve implementation
Why?
Address open comments from #11123