fix(shape): share gather and tensor contracts

[fhanuman]

Why

Gather-family and tensor-shape operations previously constructed and checked result shapes through separate converter, reifier, and verifier logic. That made dimension splicing, permutation validation, packed-int4 interpretation, and imported result-template handling easy to apply inconsistently.

This PR gives those operations shared semantic shape rules while retaining the GatherND DPS-init fallback when runtime tuple width prevents static output-rank inference.

Shape example

The added GatherBlockQuantized reification test records the physical-to-logical packed-int4 shape rule directly in HIP IR:

%result = hip.gather_block_quantized(%ctx)
  ins(%data, %indices, %scales :
      tensor<2048x96xui8>, tensor<8xi64>, tensor<2048x12xf16>)
  outs(%init : tensor<8x192xf16>)
  {bits = 4, block_size = 16, gather_axis = 0, quantize_axis = 1}
  : tensor<8x192xf16>
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%d0 = tensor.dim %result, %c0 : tensor<8x192xf16>
%d1 = tensor.dim %result, %c1 : tensor<8x192xf16>

The surviving packed axis has physical byte extent 96 and logical output extent 192; the same test suite covers the dynamic form by emitting arith.muli for that doubling.

What changes

• Share Gather and GatherND dimension-splicing rules across converters, reifiers, and verifiers.
• Define GatherElements output shape from its indices operand through the named same-shape contract.
• Validate Transpose permutations through one shared rule before materializing dimensions.
• Define Resize output N/C from the input and static spatial extents from the imported result template.
• Define Size as a rank-zero i64 result through a shared semantic rule.
• Compute GatherBlockQuantized output shapes from the logical dequantized tensor rather than physical packed-byte storage.
• Account for packed int4 by doubling a surviving quantize-axis extent, keeping converter, lowering, runtime interpretation, and tests aligned.

Stack

Full 31-PR order:

1. #688 — refactor(hip): externalize constants after ONNX conversion
2. #724 — feat(shape): preserve symbolic compiler inputs
3. #639 — refactor(hip): establish shared shape-rule foundation
4. #681 — fix(shape): canonicalize host reshape provenance
5. #725 — fix(shape): activate exact broadcasts with symbolic proofs
6. #640 — fix(hip): share MatMul and Gemm shapes with dynamic batch validation
7. #734 — fix(shape): share MatMul and Gemm contracts
8. #641 — fix(hip): unify broadcast and reduction shape contracts
9. #735 — refactor(conversion): share ONNX reduction conversion
10. #736 — fix(shape): normalize reduction contracts
11. #689 — fix(loop): support shape-changing loop carriers
12. #737 — fix(loop): support shape-changing loop carriers
13. #642 — feat(hip): add payload-driven shapes and bulk readback
14. #727 — feat(shape): consume constant payload shapes directly
15. #728 — feat(shape): add exact Tile and Range destinations
16. #729 — refactor(slice): share overflow-safe normalization
17. #731 — feat(slice): lower normalized exact descriptors
18. #733 — fix(loop): integrate exact payload Slice carriers
19. #643 — fix(hip): share Conv and Pool shape contracts
20. #738 — fix(shape): share pooling contracts
21. #742 — fix(shape): share causal convolution contracts
22. #644 — fix(hip): share Gather and tensor-shape contracts ← this PR
23. #739 — fix(shape): share block-quantized Gather contracts
24. #645 — fix(hip): share attention and normalization shape contracts
25. #740 — fix(shape): share attention contracts
26. #741 — fix(shape): share GQA capacity contracts
27. #743 — fix(gqa): reject unsupported optional features
28. #646 — refactor(hip): enforce explicit DPS shape contracts locally
29. #730 — test(hip): audit explicit DPS contract inventory
30. #732 — test(hip): audit ONNX converter deduplication
31. #647 — fix(hip): harden shape refinement failure handling

AI assistance

AI tools assisted with the shape-contract refactoring and stack reconstruction. The contributor reviewed and validated the resulting code.

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: 4473 - Commit: 323ec31

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.47	42.08	365	3	1140
GroupQueryAttention_seq128	4291.98	1.69335	12	6	209
matmul_down_seq128	520.02	2.89	69	3	255

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.55	82.92	360	3	15486

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.52	9.29	383	3	15658

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	198.0	78.6	1.23	13.59
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	334.4	40.5	1.12	6.43

OGA Wheel Smoke (Python benchmark_e2e.py)

Model	TTFT (ms)	TPS
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	191	37.2

Run: 4473 - Commit: 323ec31