fix(morphizen): guard empty output NodeArg in try_fuse

[AMDmoore]

Summary

Large hybrid/quantized models (e.g. NemotronH INT4-RTN) contain nodes with empty, non-existent NodeArg slots. Two distinct cases show up:

• Empty first output slot (the crashing case for NemotronH): a multi-output op whose first output is unused. Concretely, NemotronH's TopK consumes only its Indices output (output 1); the Values output (output 0) is unused, so output slot 0 is an empty NodeArg.
• Empty optional input slot: DequantizeLinear's optional zero_point in INT4-quantized weights.

Reading such a slot via node_arg_get_name fatal-CHECKs in node_arg.cpp (node_arg doesn't exist!), aborting the entire model compilation.

Fix:

• Add node_get_first_existing_output_name() as a node utility next to node_get_first_output_name (in node.hpp / node.cpp), and use it at every try_fuse boundary that reads a node's output name (calculate_return_values, check_loop, meta_def node emission, and Pass::fuse). It skips empty/non-existent output slots and returns the first existing output name. A fused node must still produce at least one real output, so it CHECKs (matching node_get_first_output_node_arg) rather than emitting a bogus node name.
• The isolated-node initializer collection additionally guards the empty optional input slot with node_arg_exists before reading zero_point's name.

Test plan

• Incremental build on fix branch succeeds; hipgpu.dll / hip-compiler.dll rebuilt and deployed.
• Llama-3.1-8B (INT4) via OGA run_oga_dynamic compiles and runs with no node_arg doesn't exist crash (MorphiZen EP Load ONNX Model Success + Compilation successful).
• Baseline comparison: same Llama run on main produces byte-identical output, confirming zero behavioral change for models whose output slots all exist.
• Nemotron INT4-RTN end-to-end compile (the crashing model) -- long-running (~40min), to confirm the original crash is resolved.

Made with Cursor

[github-actions]

Thanks for opening a PR!

This project follows LLVM's incremental-development and AI-tool-use
guidance. See CONTRIBUTING.md
for the project workflow.

Before requesting review, please check that:

1. The change is focused. Substantial work links the relevant issue
   or design discussion.
2. The PR documents relevant test results and updates affected
   documentation.
3. If AI tools provided substantial assistance, the description
   explains what was assisted and how it was validated, and commit
   trailers identify the tool. The contributor has reviewed and
   understands the result.

Reviewers are assigned through
CODEOWNERS where ownership
is configured.

[github-actions]

L2 Accuracy Results (EP vs CPU)

Model	Combined L2	Total Elems	Skipped NaN/Inf
conv_test_hybrid	4.8668E-07	64	0
GroupQueryAttention_seq256	25.2366	2621440	0
MatMulNBits_o_seq128	259.906	368640	0
QMoE_seq128	34.957	368640	0

Threshold: 0.01 | Run: 3493 - Commit: d48fd6d

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.51	6.29	361	3	1244
GroupQueryAttention_seq128	4381.36	1.81672	11	6	311
matmul_down_seq128	526.66	2.38	72	3	352

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.52	46.40	362	3	15590

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.52	9.63	362	3	15761

OGA Benchmark Results

Model	Warmup	Reps	Prompt Len	Gen Tokens	TTFT (ms)	TPS	Peak Mem (GB)	GPU Mem (GB)
gpt-oss-20b-webgpu-int4-rtn-block-32	1	5	128	128	161.3	80.5	1.33	13.54
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	335.8	44.1	1.22	6.43

OGA Wheel Smoke (Python benchmark_e2e.py)

Model	TTFT (ms)	TPS
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	196	43.1

Run: 3493 - Commit: d48fd6d