Profile · Projects · Fork workspaces
Status update: 2026-10-15. vLLM #59518, #58181 and DataFusion #25303 were rechecked as merged. Other entries may reflect earlier verification. This index separates authored merged pull requests, changes adopted through another upstream commit, review participation, issue reports, and open proposals. Repository names are not a substitute for the scope of a change. The linked upstream record is authoritative if a status changes after this snapshot.
PR #59518, merged 2026-10-15.
Generate preserves the engine's logprob candidates; derender applies the matching
chat or completions endpoint's top_logprobs cutoff, including the sampled-token
case. The PR reports an L20 live-server parity test reducing mismatched positions
from 215 to 0 across 1,632 comparable token positions. Tests also cover
Python/Rust implementations and sampled tokens outside the top-k set. This is a
correctness change, not an end-to-end serving speedup claim. It builds on merged
PR #58181, which added
integer token IDs to generate output logprobs.
PR #25303, merged 2026-10-14.
Extend benchmark coverage across dictionary layouts, Utf8View, plain strings,
runtime encoding, absolute distinct-value count (NDV), value width, and input batch
size, with grouped-count equivalence checks. Local results show that dictionary
encoding can help at some NDV/width combinations and regress at others. This
adds a measurement surface for future cost decisions; it does not implement an
optimizer policy or change runtime query behavior.
PR #4685, merged. The original trainer gathered full vocabulary logits on each tensor-parallel rank before calculating sampled-token logprobs and entropy. The change uses shard-local statistics and small reductions, tests logprobs and gradients against the full-vocabulary reference, and preserves the original operation sequence for batch-invariant training. The recorded 2.66× (TP2) and 3.97× (TP4) numbers are CPU/Gloo component benchmarks, not GPU measurements or end-to-end RL training speedups.
PR #25300, merged. Slice child values before floating-point zero normalization, avoiding scans and allocations over hidden backing data. In the documented Criterion case with two visible values backed by 1,048,576 Float64 values, time changes from 184.46 µs to 0.823 µs. This is a component benchmark, not a query-engine-wide acceleration. Regressions cover offsets, List/LargeList, signed zero, NaNs, and null semantics.
Merged #10411, #10413, and #10429. These changes address incomplete cache groups, redundant benchmarking after single-config pruning, and excess tensor slice indices. The contribution is a specific runtime contract and its regressions, not a claimed compiler-wide speedup.
PR #4371, merged. Oversized sandbox responses use validated continuation chunks rather than truncated JSON-RPC frames. The implementation retains host output limits and handles transfer ownership, malformed chunks, concurrent handles, and cleanup.
| Upstream PR | Change |
|---|---|
| pytorch/torchtitan #4685 | Vocabulary-parallel policy logprobs and entropy; retain the batch-invariant fallback. |
| pytorch/torchtitan #3456 | Freeze non-linear base modules during LoRA conversion. |
| Upstream PR | Change |
|---|---|
| vllm-project/vllm #59518 | Match endpoint-specific top_logprobs behavior through generate/derender; Python/Rust tests and live-server parity evidence. |
| vllm-project/vllm #58181 | Return integer token IDs for generate output logprobs; maintain Python/Rust and derender compatibility. |
| triton-lang/triton #10411 | Treat incomplete local and remote cache groups as misses. |
| triton-lang/triton #10413 | Skip benchmarking when pruning leaves one autotune configuration. |
| triton-lang/triton #10429 | Reject tensor slice indices that consume more dimensions than the source rank. |
| apache/tvm #19818 | Preserve ONNX BatchNormalization inference-mode semantics. |
| apache/tvm #19899 | Add actionable VM diagnostics for unlowered Relax operators. |
| microsoft/onnxruntime #29140 | Initialize large-head FMHA shared memory separately for each CUDA kernel variant. |
| Dao-AILab/flash-attention #2671 | Correct CuTe SM120/SM121 compile-time arguments and backward configuration. |
| ray-project/ray #64184 | Handle removed node types in autoscaler metrics without stale gauges. |
| Upstream PR | Change |
|---|---|
| apache/datafusion #25303 | Expand dictionary group-by benchmark scenarios and correctness validation; no optimizer change. |
| apache/datafusion #25300 | Normalize only visible child values in sliced floating-point list set operations. |
| apache/datafusion #23043 | Add the order-insensitive any_value SQL aggregate. |
| apache/datafusion #23066 | Make external-sort spill merge fan-in configurable. |
| apache/datafusion #23226 | Decode Hive partition values while preserving the borrowed fast path. |
| apache/datafusion #23232 | Align scalar UDF literal arguments with their coerced return fields. |
| apache/arrow-rs #10298 | Preserve LargeList schema types through Flight encoding. |
| apache/arrow-rs #10907 | Check remaining temporal-cast overflows and preserve safe-cast null semantics. |
| Upstream PR | Change |
|---|---|
| UKGovernmentBEIS/inspect_ai #4371 | Preserve oversized sandbox JSON-RPC responses through bounded continuation. |
| scikit-learn/scikit-learn #34380 | Avoid the negative-stride NumPy-to-PyTorch DLPack process abort. |
| mlc-ai/xgrammar #667 | Handle negative-zero decimal ranges without broadening numeric boundaries. |
| carla-simulator/carla #9791 | Restore the LiDAR smoke-helper instance-method signature. |
SWE-bench skipped-test grading: PR #598
is closed and is not a merged PR. The change was adopted through
upstream commit a5ecda6640d13f89848a3dceaa08585431d258db,
whose message references #598. Skipped FAIL_TO_PASS tests must not count as a
successful repair; skipped PASS_TO_PASS tests retain their separate maintenance semantics.
| Record | Contribution type |
|---|---|
| DataFusion #24890 | Review contribution; not an authored implementation PR. |
| Gradio #13556 → upstream fix #13580 | Issue report and linked upstream remediation; the fix is not counted as my authored PR. |
| PyTorch #188023 | Negative-stride DLPack process-abort report; separate from the merged scikit-learn workaround. |
These are proposals, not shipped upstream features. A local test, source overlay, or component benchmark is not presented as completed upstream integration.
| Proposal | Technical scope | Evidence boundary |
|---|---|---|
| vLLM #57442 | Sampled-token logprob transport without the general top-k object path | L20 serving A/B; exactness checks use batch-invariant execution. Open. |
| vLLM #58198 | Request-free CPU→GPU KV prefetch with explicit submission/completion states | Reserve-gated primitive. Benefits under slack do not establish gains under occupancy. Open. |
| Arrow #11062 | Compression-preserving RunEndEncoded numeric kernels | Differential tests and component benchmarks; low-compression cases can lose. Open. |
| DataFusion #23065 | Native grouped any_value accumulation | Component-level benchmarks, not whole-query gains. Open. |
| SGLang #41498 | Fit FP16 batch-invariant GEMM within the device's shared-memory limit | Actual RTX 4090/5090 checks; resource-correctness fix, not an infinite-speedup claim against a crashing baseline. Draft/open. |
| vLLM #58953 | Avoid an unused host-length synchronization for non-causal XQA decode | Actual CUDA checks on the recorded source-overlay setup; current-main integration remains outstanding. Draft/open. |
All authored PR states · Merged PR search · Open PR search
Search visibility can depend on GitHub sign-in. The direct links above remain the curated public evidence; this page is not a claim that every historical issue or review has been enumerated. Own-repository and affiliated-project maintenance is not counted as independent upstream adoption.
All contribution forks and their branch histories are retained. Upstream READMEs, licenses, genuine assistance disclosures, and active PR branches are not rewritten for profile presentation. See the workspace index.