정확성을 검증하는 AI/ML 시스템 엔지니어

[inductor] remove_noop_ops: keep a copy of an input that is mutated before the copy's user runs

[inductor] Realize consumers of aliasing buffers before a mutation of the aliased buffer

[dynamo] Graph break on an in-place view of a no-op alias of a graph input

[inductor] Do not reinplace into a graph input that a later, unrelated mutation overwrites

[aot_autograd] Regenerate output aliases of merged (synthetic-base) inputs from metadata

[inductor] `dst[:] = src[:]; src.add_(1)` copies the *updated* `src` into `dst` (remove_noop_ops replaces a copy of an input by the input itself, which is mutated before the copy's user runs)

[inductor] Consumer of a dtype view of an input is computed after a later in-place mutation of that input (y = x.view(int32) * 2; y.sub_(-4); x[:, 2:5] = 2; return y.view(int64))

[functionalization] In-place mutation through a gapped unfold view (step > size) silently zeroes every base element not covered by a window under torch.compile

[dynamo] x.contiguous().unsqueeze_(d) / x.to(x.dtype).unsqueeze_(d) crashes guard creation with IndexError: list index out of range (in-place view through a no-op alias of the input)

[aot_autograd] Output that is a view of an aliased input (synthetic base) is regenerated from the wrong tensor: ta[0] returns t0[0], ta.view(-1) asserts on ViewMeta shape

[inductor] Scatter result reinplaced into a graph input that a later, unrelated mutation overwrites: y = slice_scatter(x, src); x.zero_(); return y returns zeros

[functionalization] unfold inverse: keep base elements that no window covers

[Fix][Relax] Accept a returned tuple bound to a Var in Gradient

[FIX][Relax] FuseTransposeMatmul: skip when the transpose has other users

[FIX][LLVM] Support bool Add/Sub/Mul in LLVM codegen

Dynamic VRAM: a ModelSamplingDiscrete schedule leaks into later runs of the same checkpoint (buffers restored by attribute path), giving wrong or black images reported as success

[Bug] `--json-model-override-args '{"rope_scaling": ...}'` drops `rope_theta` for models without a per-model fallback: Llama-3.2-3B-Instruct GSM8K 161 → 106 (first 200)

• 다른 개발자의 수정 PR SGLang#41239 ↗ ○ 닫힘
• 다른 개발자의 수정 PR SGLang#41327 ↗ ○ 닫힘

[Bug]: `--hf-overrides '{"rope_scaling": ...}'` drops `rope_theta` under Transformers v5; models without a per-file default silently run with base 10000

• 다른 개발자의 수정 PR vLLM#58679 ↗ ● 리뷰 대기