Fix Cast node naming collisions and opset 10 Resize in float16 conversion

[Rishi-Dave]

Summary

• Fix Cast node naming collisions in convert_float_to_float16 when nodes have empty names (common in PyTorch exports)
• Fix ALWAYS_FLOAT_INPUTS for opset 10 Resize where scales at index 1 was unprotected
• Add dedicated test suite for float16 conversion (test_float16.py, 8 tests)

Motivation

Fixes #14827

When convert_float_to_float16 processes models with unnamed nodes (empty node.name, very common in PyTorch/TensorFlow-exported ONNX models), the generated Cast node names collide. For example, multiple Resize nodes all produce Cast nodes named "_input_cast_2" and output tensors named "_input_cast_2", corrupting the graph with duplicate names.

Additionally, the ALWAYS_FLOAT_INPUTS dict only protected Resize scales at index 2 (opset 11+ layout: [X, roi, scales, sizes]), but opset 10 Resize has scales at index 1 ([X, scales]), leaving it unprotected.

Changes

onnxruntime/python/tools/transformers/float16.py (11 lines changed):

• Use unique tensor names (input_name/output) as the base for generated Cast node and output names, instead of potentially-empty node.name
• Add index 1 to ALWAYS_FLOAT_INPUTS["Resize"] to protect opset 10 scales
• Fix misleading comment ("change current node's input name" → "output name")

onnxruntime/test/python/transformers/test_float16.py (new file, 8 tests):

• test_resize_opset11_cast_naming_unique — multiple unnamed Resize nodes produce unique Cast names
• test_resize_opset11_scales_initializer_stays_fp32 — scales initializer preserved as float32
• test_resize_opset10_scales_initializer_stays_fp32 — opset 10 scales protected at index 1
• test_resize_opset10_multiple_unnamed_unique_names — opset 10 naming uniqueness
• test_blocked_node_cast_naming_unique — blocked op nodes (Upsample) also get unique Cast names
• test_resize_with_op_block_list — Resize in op_block_list still produces unique names
• test_data_input_converted_to_fp16 — data tensor correctly converts to fp16
• test_force_fp16_initializers — force flag overrides protection

Test Plan

• All 8 new tests pass locally (python -m unittest test_float16.TestFloat16Conversion -v)
• Existing test_gpt2_past_fp16 test passes (no regression in existing float16 behavior)
• ruff check passes on both files

[Rishi-Dave]

@microsoft-github-policy-service agree

[tianleiwu]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[Copilot]

Pull request overview

This pull request fixes critical issues in the float16 conversion tool that caused graph corruption when processing models with unnamed nodes (common in PyTorch/TensorFlow exports) and incorrectly converted Resize scales in opset 10 models.

Changes:

• Fixed Cast node naming to use unique tensor names instead of potentially-empty node names, preventing naming collisions
• Added protection for Resize scales at input index 1 (opset 10 compatibility)
• Added comprehensive test suite with 8 tests covering naming uniqueness, opset 10/11 Resize handling, and edge cases

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/python/tools/transformers/float16.py	Updated ALWAYS_FLOAT_INPUTS to protect Resize index 1; changed Cast node naming from node.name + "_input_cast_" + str(i) to input_name + "_cast_to_fp32"; fixed misleading comment
onnxruntime/test/python/transformers/test_float16.py	New test file with 8 tests covering naming uniqueness for unnamed nodes, opset 10/11 Resize scales protection, blocked ops, and force_fp16_initializers behavior

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tianleiwu]

The core fix (using tensor names instead of potentially-empty node.name for Cast node naming) is correct and well-tested.

One concern:

Opset 11+ roi over-protection: Adding index 1 to ALWAYS_FLOAT_INPUTS["Resize"] correctly protects opset 10 scales, but also forces opset 11+ roi to stay fp32 — even though the ONNX spec allows roi to be fp16 (T2 constraint). Practically harmless since roi is usually empty, but technically imprecise. Suggest at minimum adding a comment, or ideally making protection opset-aware.

Other minor notes: shared-input edge case not fully addressed (not a regression), and the naming convention change could theoretically affect downstream tooling that pattern-matches Cast node names.

[Rishi-Dave]

Thanks for the thorough review @tianleiwu! Great catch on the opset 11+ roi over-protection.

I've pushed a fix that makes the protection opset-aware:

• Reverted ALWAYS_FLOAT_INPUTS["Resize"] back to [2] (opset 11+ default)
• Added logic inside convert_float_to_float16 that detects the ONNX opset version from model.opset_import and adjusts at runtime:
  • Opset ≤ 10 ([X, scales]): protects index 1 (scales)
  • Opset 11+ ([X, roi, scales, sizes]): protects index 2 (scales) only; roi at index 1 is left to fp16 per the ONNX spec

Updated the test to verify that opset 11+ roi is not over-protected.

Re: the shared-input edge case and naming convention — agreed these are pre-existing behaviors, not regressions from this PR. Happy to address them in a follow-up if you'd like.

[tianleiwu]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).