Bot review (gpt-5.6-sol) — DM the bot to share feedback.

The fix is straightforward and consistent with existing exporter patterns: GraphSurgeon topologically sorts the graph after the NVFP4 post-processing code appends producer nodes, and the new CPU regression test exercises the real Torch→ONNX→NVFP4 conversion path and validates the result with ONNX checker. Imports are correctly placed, the existing ONNX optional dependency already includes GraphSurgeon, and no licensing changes are introduced.

Claude review passed — no blocking issues found. LGTM

Scope: The PR's authoritative changed-file list is 2 files — modelopt/onnx/export/nvfp4_exporter.py and tests/unit/torch/quantization/test_onnx_export_cpu.py. (Local base drift also surfaced trt_utils.py/test_plugin.py hunks, but those are not part of this PR and were excluded.) Both PR files were reviewed in full.

Findings — CRITICAL: 0, IMPORTANT: 0, SUGGESTION: 0

Assessment (low risk):

• The fix appends a GraphSurgeon import_onnx → toposort() → export_onnx round-trip at the end of NVFP4QuantExporter.post_process(), correctly reordering the newly-inserted DequantizeLinear/Cast producer nodes ahead of the Transpose/MatMul consumers they feed. This matches the established pattern used by the sibling FP8/INT4/INT8 exporters and graph_utils, and onnx-graphsurgeon>=0.6.1 is already a declared dependency (no new import risk).
• The sole caller (_deploy/utils/torch_onnx.py:477) reassigns process_model()'s return, so returning a fresh ModelProto from gs.export_onnx() is safe — nothing relies on in-place identity.
• The new CPU regression test drives the real Torch→ONNX→NVFP4 path, asserts TRT_FP4QDQ is present pre-conversion and gone post-conversion, and validates with onnx.checker.check_model (which itself enforces topological ordering) — a real guard for this bug.

Actionable comments posted: 1

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@modelopt/onnx/export/nvfp4_exporter.py`:
- Line 467: Update _cast_input_dtypes() to cache and reuse Cast outputs keyed by
(input_name, precision_dtype), ensuring shared activations produce only one Cast
node and tensor. Preserve the existing <input>_f16 naming where applicable, and
ensure _topologically_sort_graph_nodes() receives a graph without duplicate
producers.

---

Nitpick comments:
In `@modelopt/onnx/export/nvfp4_exporter.py`:
- Line 467: Extend the regression test covering the export flow around
_topologically_sort_graph_nodes to assert that every in-graph producer node
appears before its consumer, preferably using a branched graph. Keep the
existing FP4QDQ-removal and onnx.checker.check_model checks, and make the new
assertion validate the sorted node order directly.
- Around line 62-71: Replace the list-based ready queue in the topological
sorting flow with a heapq min-heap keyed by original node indices. Heapify the
initial ready indices, use heap push/pop operations when nodes become ready, and
remove the per-iteration sorting while preserving stable ascending node ordering
and the existing dependency processing.

Actionable comments posted: 1

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@modelopt/onnx/utils.py`:
- Around line 59-67: Update topologically_sort_graph_nodes to account for
tensors captured by nested If/Loop subgraphs when building dependencies, either
by recursively collecting those outer-scope references or by explicitly
rejecting graphs containing unsupported control-flow nodes. Preserve direct
input dependency handling and add a regression test covering a producer and
control-flow consumer connected through an implicit capture.

---

Nitpick comments:
In `@modelopt/onnx/utils.py`:
- Around line 69-78: Replace the ready-list queue in the topological sorting
logic with a heapq min-heap, updating initialization, extraction, and insertion
accordingly. Preserve ascending node-index ordering and the existing dependency
processing in the loop, while removing repeated pop(0) and ready.sort()
operations.

Bot review (gpt-5.6-sol) — DM the bot to share feedback.

Re-review complete. The prior correctness concerns are addressed: shared activation Casts are now cached per (input_name, precision_dtype) with regression coverage, and the protobuf topological sorter accounts for outer-scope tensors captured by nested control-flow subgraphs with a focused test. The exporter regression still exercises the real Torch→ONNX→NVFP4 path and validates the converted graph with ONNX checker. The remaining heap-based ready-queue suggestion is a non-blocking performance nit; the current stable sort is correct. No licensing changes or blocking issues found.

Claude review passed — no blocking issues found. LGTM

Scope: Authoritative PR file list is 3 files — modelopt/onnx/export/nvfp4_exporter.py, modelopt/onnx/utils.py, and tests/unit/torch/quantization/test_onnx_export_cpu.py. All three reviewed in full. (trt_utils.py/test_plugin.py hunks in the local diff are shallow-checkout base drift, not part of this PR — excluded.)

Findings — CRITICAL: 0, IMPORTANT: 0, SUGGESTION: 0

Assessment (low risk):

• topologically_sort_graph_nodes implements Kahn's algorithm correctly: duplicate-producer / self-dependency / cycle guards raise instead of silently misordering; graph inputs & initializers correctly seed the ready set; stable ascending-index ordering is preserved. Outer-scope tensors captured by If/Loop subgraphs are folded into the dependency set with correct recursive propagation, and the new regression test exercises the producer-after-consumer capture case.
• Shared-activation Cast reuse is keyed by (input_name, precision_dtype); since precision is computed once per graph this dedupes by input, so multiple NVFP4-lowered MatMuls sharing an activation reuse a single Cast. The _f16/_bf16 suffix is an internal tensor name only — nothing downstream string-matches these names, so it is behavior-safe.
• The sort runs in place after initializer cleanup and post_process returns the same ModelProto; no mode registration, modelopt_state schema, or public export-config surface is touched. Pure ONNX graph-ordering bug fix.
• The only open reviewer item is CodeRabbit's non-blocking heapq micro-optimization; export is not a hot path and the current stable sort is correct, so it does not block.