fix(gbq): key GatherBlockQuantized default zero point off ONNX storage type, not signedness

[BoarQing]

Summary

GatherBlockQuantized's default zero point (used when zero_points is absent) depends on the ONNX storage type T1, which the EP currently cannot observe. tensor(uint4) and tensor(uint8)-holding-packed-nibbles both arrive at the runtime as ui8 with bits=4 and a byte-identical layout, yet the ORT CPU kernel uses 0 for the former and 2^(bits-1) for the latter:

// onnxruntime/contrib_ops/cpu/quantization/gather_block_quantized.cc
if constexpr (std::is_same_v<T1, uint8_t>) {
  ...
  const int32_t default_zero_point = 1 << (static_cast<int>(bits_) - 1);
} else {  // Int4x2 / UInt4x2
  zp_val = zero_points_ptr ? ... : 0;
}

Deciding from is_signed_data, as main does today, therefore biases one of the two by 8 * scale whichever way the condition is written. In practice uint4 embedding tables published without zero_points come back shifted by -8 * scale on every gathered row.

The correct rule is already written down in three places in this repo — HipOps.td ("the default is 0 for int4/uint4 and 2^(bits-1) for uint8", from #368), hipdnn_ep_runtime.h, and the ORT schema doc string — but is not what the code implements.

Approach

Carry the ONNX element type explicitly rather than trying to re-derive it:

1. Import (mlir-graph.cpp) records onnx.element_type on every onnx.Constant. This attribute was already listed in the GBQ legalizer's preserve-set but nothing ever set it.
2. convert-onnx-to-hip pins quant_storage_bits from that marker, before legalizeInt4ConstantIfNeeded rewrites the constant and erases the shape evidence that would otherwise distinguish the two cases.
3. convert-hip-to-llvm forwards it to the runtime (22nd parameter).
4. Runtime selects default_zp = (storage_bits == 8) ? 1 << (bits - 1) : 0.

Signedness no longer participates in the zero-point decision at all. It is still needed for nibble sign extension, which is a separate concern — the previous code used one boolean for both jobs and got the second one wrong.

quant_storage_bits is declared in ODS with a verifier rather than attached as a discardable attribute, so producer and consumer share a generated accessor and illegal widths are rejected at verification. It is optional and falls back to bits, keeping hand-written IR valid.

Also routes the prepare-pattern attribute edits through rewriter.modifyOpInPlace instead of mutating the op behind the driver's back.

Verification

Built main and this branch separately and ran both against the same models with the same ORT CPU reference (hip-onnx-runner -L, EP vs -n CPU):

Model	Storage / bits	zero_points	main L2	this PR L2
uint4_nozp	uint4 / 4	no	166.277	0
uint4_nozp_100352x2048	uint4 / 4, external	no	940.604	0
int4_nozp	int4 / 4	no	0	0
uint8_nozp	uint8 / 8	no	0	0
u8bits4_nozp	uint8 / 4 packed	no	0	0
uint4_zp	uint4 / 4	yes	0	0
uint4_zp_blk16	uint4 / 4, block 16	yes	0	0
int4_zp	int4 / 4	yes	0	0
uint8_zp	uint8 / 8	yes	0	0
uint4_zp_100352x2048	uint4 / 4, external	yes	0	0

The failure signature on main is a pure additive shift, not noise or a layout error:

ORT CPU (truth) : [0.   0.5  1.   1.5  2.   2.5  3.   3.5]
main            : [-4.  -3.5 -3.  -2.5 -2.  -1.5 -1.  -0.5]
diff            : [-4.  -4.  -4.  -4.  -4.  -4.  -4.  -4.]   == -8 * scale

u8bits4_nozp is the case that separates the two candidate rules: it is byte-identical to uint4_nozp but requires the opposite default. It passes on both builds, confirming the fix keys off storage type and not off bits.

361/361 lit tests pass. ORCA-2bit runs end to end; it is unaffected either way because it ships zero_points, so the default is never consulted.

Affected models

Only when all three hold: storage is tensor(uint4), bits == 4, and zero_points is absent. Anything shipping zero_points is unaffected.

Not addressed here

Deliberately out of scope, flagged for follow-up:

• bits == 2 is legal for tensor(uint8) per the schema, but convert-onnx-to-hip still gates on bits in {4, 8}, so such models fall back rather than converting.
• quantize_axis is still re-derived from shape invariants when the attribute is absent, rather than using the ONNX default of 1.
• legalizeInt4ConstantIfNeeded and recreateExternalConstant still return false / nullptr silently instead of emitting diagnostics.
• The CI L2 suite still contains no GBQ model. This is the gap that let a wrong default zero point ship. The 10 models above are currently local only; wiring them (or at least uint4_nozp and u8bits4_nozp) into the gate is the change that prevents a recurrence.

Test plan

• CI L2 suite green
• lit suite green
• Reviewer sanity-check the storage-width rule against the ORT CPU kernel
• Decide whether the GBQ models should be added to the CI L2 gate in this PR or a follow-up

[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: 3486 - Commit: 143c44e

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.51	6.10	365	3	1244
GroupQueryAttention_seq128	4384.48	1.80928	11	6	311
matmul_down_seq128	521.26	2.39	75	3	351

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.52	45.87	362	3	15590

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.50	9.89	359	3	15760

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	173.1	78.9	1.33	13.54
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	375.7	35.3	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	195	35.5

Run: 3486 - Commit: 143c44e