Fix/pinned portable flags#115
Open
felipeblazing wants to merge 5 commits intoNVIDIA:mainfrom
Open
Conversation
felipeblazing
force-pushed
the
fix/pinned-portable-flags
branch
from
b753c0f to
62e0517
Compare
…pool peer access Squash of 5 fixes onto cucascade 73d00c4 for Sirius multi-GPU v1.4. Original commits: 1fff85d 3743621 2dcab24 ff14ff4 e23f3a2 - Pin host memory with cudaHostAllocPortable for multi-GPU DMA accessibility - Make cudaMallocHost sites Portable-aware - Add cudaHostAllocMapped to all pinned allocation sites - Per-instance ptds_allocation_tracker (thread_local, not process-global) - Drop pool priming (prevents multi-space GPU memory exhaustion) - Add cross-device pool peer access at construction time Co-Authored-By: Felipe Aramburu <faramburu@nvidia.com>
Squash of 1 fix onto cucascade 73d00c4 for Sirius multi-GPU v1.4. Original commit: eda349a std::thread _thread must be declared AFTER _mutex and _cv. C++ destroys members in reverse-declaration order; placing _thread last ensures the join() inside ~io_worker happens while _mutex and _cv are still alive. Avoids EINVAL on mutex destruction during io_worker teardown under parallel test runs. Co-Authored-By: Felipe Aramburu <faramburu@nvidia.com>
…A probe at init Squash of 3 fixes onto cucascade 73d00c4 for Sirius multi-GPU v1.4. Original commits: 7ed84f2 cc2a53d e4db3d8 - Use target-bound stream in host->gpu and gpu->gpu converters (v1.1 P2P fix) - Pass source mr to cudf::pack + default-pool peer access (WIP carry) - Empirical P2P peer DMA probe at init; route convert_gpu_to_gpu to real peer DMA on server hardware and host-staging on consumer chipsets Co-Authored-By: Felipe Aramburu <faramburu@nvidia.com>
…tation + cudaStreamWaitEvent Squash of 2 fixes onto cucascade 73d00c4 for Sirius multi-GPU v1.4. Original commits: 7409c60 62e0517 Phase 13 fix: closes SF100 Q11 2-GPU illegal-address race. - Add record_writer_event/get_writer_event accessors on gpu_table_representation - require writer_stream as a REQUIRED ctor argument (compile-time enforced) - convert_gpu_to_gpu: cudaStreamWaitEvent(target_stream, src.get_writer_event()) before peer copy; fallback to cudaDeviceSynchronize for un-migrated callers - Replaces cudf::pack path with column-tree walk (avoids stream-ordered race in compute_splits scratch allocations) Co-Authored-By: Felipe Aramburu <faramburu@nvidia.com>
Pure formatting and typo fixes from `pre-commit run --all-files`: - 15 source/header files reformatted by clang-format (line-wrap, brace alignment) — no semantic changes - docs/ARCHITECTURE.md: 'sytem' → 'system' (codespell) No semantic changes; runtime behavior identical to v1.4 ship state (cucascade pin 1c1e648). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
felipeblazing
force-pushed
the
fix/pinned-portable-flags
branch
from
62e0517 to
42a01c4
Compare
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
1 participant
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.
No description provided.