Peter Chen

Replaced training-path batch variance `E[x^2] - E[x]^2` with centered two-pass `E[(x - mean)^2]` for bf16 inputs with large channel offsets and low variance, preventing cancellation-driven NaNs or badly scaled normalization while leaving inference and device kernels unchanged. Link
Removed a width-only free-L1 heuristic from `untilize_with_unpadding` operation attributes after tracing program-factory selection: the heuristic changed graph-capture and program-cache identity without changing the selected factory. Added cache-count and captured-argument regressions while retaining the height heuristic that does select a factory. Link
Balanced the lifecycle of local partial-reduction circular-buffer pages in four sharded LayerNorm multicast kernels by pairing `wait_front` / `get_read_ptr` consumption with `pop_front` after gather-combine, without changing semaphore ordering, NOC topology, or reduction order. Link
Tightened the `group_norm` launch contract by validating user-provided `input_mask` and `negative_mask` before `ttnn::device_operation::launch(...)`, converting host-backed optional-mask faults into operator-owned errors without changing sharding math or program factories. Link
Fixed rank-1 `moreh_norm` and `moreh_sum` global-reduction edge cases by deriving reduction expansion from logical rank, preserving scalar-output semantics through the existing width-reduction path instead of indexing padded-shape coordinates as logical dimensions. Link Link

Specialized `DenseMatrix::vertically_packed_row` and `vertically_packed_row_pair` for the scalar no-packing case `P::WIDTH == 1`, replacing generic wraparound row-slice materialization with direct contiguous loads while keeping packed-width semantics on the original path. Criterion measurements showed `-45.26%` to `-83.56%` latency on the target scalar benchmarks; packed-path deltas remained effectively flat (`-0.74%` to `+1.34%`).

Link

Removed redundant `array.new` / `array.write` / `array.read` carrier traffic in static-`K欠 struct-of-pods lowering: when all reader `K欠 values are compile-time constants and writer reads dominate their uses, bind the hoisted `struct.readm` values directly while retaining the carrier fallback for runtime-`K欠 and hoist-bridging cases. Link
Fixed mixed batched plus rank-0 constant-index `stablehlo.dynamic_update_slice` lowering by adding constant-index axes to the one-hot scatter mask, preventing an unconstrained axis from broadcasting a write across an entire dimension. Link
Replaced incidental printed-text parsing with structured MLIR API queries for initialized `pod.new` records, pod-typed array legality, and `ool.cmp` predicates, so lowering decisions are resilient to printer spelling changes. Link

Instrumented `MSM::batch_multi_scalar_mul(...)` with phase-level `BB_BENCH欠 scopes for work-unit evaluation, result accumulation, batch normalization, and Montgomery conversion; added `BatchMSM_1656` for `2^16` and `2^20` point MSMs at `256` threads, turning a suspected final-reduction bottleneck into a measurable claim. Link

Converted print-only historical `nlsat` observations into structural assertions, including a projected-lemma counterexample guard that fails when the historical unsoundness shape reappears instead of silently printing diagnostic state. Link

Separated public rotation-gate filter identity from internal generator normalization in `CommutationChecker`, so a filter for `rx` accepts `RXGate` pairs while the commutation engine can still reason over the `x` generator; added positive, mixed-filter, and rejected-filter regressions. Link

Added packet-version-number validation to `SpacePacketDeframer` before downstream forwarding, preserving valid command, secondary-header, and sequence-flag variants while routing malformed PVN packets through the existing error-notify and buffer-return path. Link

Preserved multi-device tensor placements in graph-report import by changing deduplication from `(tensor_id, address)` to `(tensor_id, device_id, address)` and pinning both collision preservation and exact-duplicate collapse in importer-only regression tests. Link
Added pre-launch validation for `moreh_group_norm_backward` and `moreh_layer_norm` stats tensors so invalid shape, device, or tensor contracts fail at the operator boundary rather than inside execution setup. Link Link
Stabilized `test_sum_global` golden behavior across host OpenMP settings, isolating device reduction semantics from host parallel-runtime variance. Link

Fixed two-phase `eth_estimateGas` fallback assembly so an actual top-level revert is returned as `ExecutionReverted` instead of being masked by an earlier intrinsic-gas probe failure, while retaining allowance and balance-error priority. Link
Added strict common-converter validation for EIP-1474 balance inputs and locked engine task cleanup, proof-response, zero-gas, and block-selector boundary semantics with focused regression coverage. Link Link Link

Batched consecutive token-bridge outputs in the Rust/CUDA `vllm-frontend` streaming path, eliminating avoidable per-output bridge overhead between frontend orchestration and runtime delivery. Link

Removed a Pytest collection warning in the Gluon frontend without weakening coverage by renaming a `@gluon.jit` parser helper whose `test_` prefix caused the JIT wrapper to be collected as a test candidate; verified under warnings-as-errors. Link

Bounded stalled driver fetches with a local 10-second timeout while preserving producer shutdown and SQL semantics; added non-blocking warning observability for job and queue subscription drops caused by full subscriber buffers. Link Link

Stabilized REST and MCP `listScoreConfigs()` pagination with a total order `createdAt DESC, id ASC欠, preventing overlaps and skipped entries when timestamps tie. Link