Remove Hugging Face download dependency from test_dynamic_quantization_subgraph

[Copilot]

Description

test_dynamic_quantization_subgraph downloaded decoder_model_merged.onnx from huggingface.co at test time and only caught urllib.error.HTTPError. DNS/connection failures raise the broader urllib.error.URLError, which escaped and failed CI (e.g. macOS job 99624951799, run 33433704903).

No trustworthy Microsoft/ONNX-Runtime-controlled GitHub copy of the exact unquantized model exists, so instead of adding a broader exception handler around a network call, the test now builds a minimal ONNX fixture locally, removing the network dependency entirely.

• Local fixture instead of download: _make_merged_decoder_like_model() constructs a small model with a top-level float initializer shared.weight_merged_0 consumed by a Gather node inside each branch (then_branch/else_branch) of an If node — reproducing the structural property under test (outer-scope weight sharing across nested subgraphs), matching the real merged T5 decoder's with-past/without-past branches.
• Same assertions preserved: quantizing with extra_options={"EnableSubgraph": True} still must hoist shared.weight_merged_0_quantized / shared.weight_merged_0_scale to the top-level graph and must not leak shared.weight* initializers into subgraphs.
• No pre-quantized substitute: the fixture is unquantized float data, so the test still genuinely exercises quantization behavior rather than short-circuiting it.

gather_node = onnx.helper.make_node("Gather", [shared_weight_name, "input_ids"], [gather_output], axis=0)
# ... wrapped in then_branch/else_branch of an If node, with shared_weight_name
# defined only as a top-level initializer.

Motivation and Context

Eliminates a flaky, network-dependent CI failure and makes the test deterministic, while still covering the intended subgraph-quantization initializer-hoisting behavior.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟢 Approval recommended

The deterministic fixture preserves the intended coverage, with no unresolved issues.

Pull request overview

Replaces a flaky Hugging Face model download with a deterministic local ONNX fixture while preserving subgraph quantization coverage.

Changes:

• Builds a merged-decoder-like If model locally.
• Removes the external network dependency.
• Retains initializer-hoisting assertions.

File summaries

File	Description
onnxruntime/test/python/quantization/test_subgraph.py	Adds the local fixture and uses it in the quantization test.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Justin Chu (justinchuby)]

FYI you may also use the onnx textual format to simplify the fixture code