Fuse DQ->MatMulNBits for shared initializers and skip outer-scope weights - #33800

Description

DQMatMulNodeGroupSelector accepts a DequantizeLinear whose weight, scale and zero point are initializers of an enclosing graph (GetConstantInitializer(name, true)), but TransposeDQWeightsForMatMulNBits reads them with Graph::GetInitializedTensor, which only searches the current graph. So a DQ -> MatMul inside an If/Loop body whose weights live in the outer graph fails session creation:

FAIL : qdq_actions.cc:132 TransposeDQWeightsForMatMulNBits Missing required weight: w_q for node: 

This is what #28306 hit: in onnx-community/whisper-tiny decoder_model_merged_uint8.onnx the embedding scale lives in the main graph and both DQ nodes are inside the optimum::if branches. It still reproduces on 1.31.1 with a single unshared DQ -> MatMul in an If branch (repro below), because the guard added in #28326 only looks at consumer counts.

That guard (reject when the weight or scale initializer has more than one consumer) also turns the fusion off for every flat tied-embedding model, where the embedding Gather and the lm_head share weight, scale and zero point. Those models dequantize the whole lm_head weight on every decode step. For HuggingFaceTB/SmolLM2-135M-Instruct model_int8.onnx that is a [576, 49152] uint8 -> fp32 DequantizeLinear per token, 44% of int8 decode time on a Neoverse N2.

This change:

• requires the weight/scale/zero point to be initializers of the node's own graph (GetConstantInitializer(name, false)), matching what the action can read, and
• removes the shared-initializer guard. The action only reads the originals and adds new transposed initializers, so sharing is safe; the originals are cleaned up once unused.

Tests: the two *_SharedWeight tests and the A16 shared case now expect fusion (outputs still compared against the unfused graph), and DQMatMulNotConvertedToMatMulNBits_OuterScopeWeight covers DQ -> MatMul inside If branches with outer-scope weights.

Motivation and Context

Decode, ms/token, CPU EP, GitHub ubuntu-24.04-arm (Azure Cobalt 100 / Neoverse N2, 4 threads), onnx-community / HuggingFaceTB model.onnx vs model_int8.onnx:

Same models on GitHub ubuntu-24.04 (AMD EPYC 7763, 2 threads):

Accuracy, against 1.31.1 on the same int8 models:

• default session.qdq_matmulnbits_accuracy_level (4): lm_head logits differ by at most 1.6% relative; WikiText-2 perplexity 35.277 -> 35.345 (SmolLM2) and 63.199 -> 63.269 (Qwen).
• session.qdq_matmulnbits_accuracy_level=1: logits are bit-identical to 1.31.1. Note this path is slow for 8-bit weights today, because there is no SQ8BitGemmVariant_CompFp32 and MatMulNBits<float>::ComputeBUnpacked dequantizes B on every call (the existing TODO(fajin): move B dequant to prepack). That already applies to any 8-bit DQ -> MatMul that gets fused; with this change it also applies to tied lm_heads. Happy to follow up on that path separately.

Both whisper-tiny decoder_model_merged_uint8.onnx and decoder_model_merged_quantized.onnx load, with the same ops as 1.31.1.

Repro for the outer-scope crash on 1.31.1:

import numpy as np, onnx, onnxruntime as ort
from onnx import helper as h, TensorProto as T, numpy_helper as nh
K, N = 64, 32
def branch(name):
    dq = h.make_node("DequantizeLinear", ["w_q", "w_scale", "w_zp"], [f"w_{name}"])
    mm = h.make_node("MatMul", ["x", f"w_{name}"], [f"y_{name}"])
    return h.make_graph([dq, mm], name, [], [h.make_tensor_value_info(f"y_{name}", T.FLOAT, None)])
inits = [nh.from_array(np.zeros((K, N), np.uint8), "w_q"), nh.from_array(np.array(0.02, np.float32), "w_scale"),
         nh.from_array(np.array(3, np.uint8), "w_zp")]
g = h.make_graph([h.make_node("If", ["cond"], ["y"], then_branch=branch("then"), else_branch=branch("else"))], "g",
                 [h.make_tensor_value_info("cond", T.BOOL, []), h.make_tensor_value_info("x", T.FLOAT, [1, K])],
                 [h.make_tensor_value_info("y", T.FLOAT, None)], inits)
onnx.save(h.make_model(g, opset_imports=[h.make_opsetid("", 21)], ir_version=10), "m.onnx")
ort.InferenceSession("m.onnx")  # 1.31.1: Missing required weight: w_q

Fixes #28306 properly (follow-up to #28326). cc Tianlei Wu (@tianleiwu)