Reuse quantized embedding table for tied LM head in TieWordEmbeddings#2549

Open

justinchuby wants to merge 13 commits into

mainfrom

justinchu/tie-embeddings-reuse-quantized

Commits on Jul 1, 2026

Add RewriteRuleSurgeon base and port first graph surgeries to onnx_ir

justinchuby
and
Copilot
committed
Port RemoveGidxFromMatMulNBits to onnx_ir

justinchuby
and
Copilot
committed
Port InferShapes, RemoveShapes, RemoveInputs to onnx_ir

justinchuby
and
Copilot
committed
Port ZeroOutInput to onnx_ir

justinchuby
and
Copilot
committed
Port RemoveMemcpy to onnx_ir

justinchuby
and
Copilot
committed
Port ReplaceAttentionMaskValue to onnx_ir

justinchuby
and
Copilot
committed
Port RMSNorm/LayerNorm surgeries to onnx_ir

justinchuby
and
Copilot
committed
Port MatMulAddToGemm and GemmToMatMulAdd to onnx_ir

justinchuby
and
Copilot
committed
Port GQA/quant graph surgeries to onnx_ir

justinchuby
and
Copilot
committed
Port QuantizeEmbeddingInt8 and ShareEmbeddingLmHead to onnx_ir

justinchuby
and
Copilot
committed
Add tests for PowReduceSumPowDiv2LpNorm and QuantizeEmbeddingInt8

justinchuby
and
Copilot
committed
Address PR review comments on graph surgery migration

justinchuby
and
Copilot
committed
Reuse quantized embedding table for tied LM head in TieWordEmbeddings

justinchuby
and
Copilot
committed