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

Open

Justin Chu (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

Justin Chu (justinchuby)
and
Copilot
committed
49c5096
Port RemoveGidxFromMatMulNBits to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
b0053eb
Port InferShapes, RemoveShapes, RemoveInputs to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
574e766
Port ZeroOutInput to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
6be8173
Port RemoveMemcpy to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
997fe4e
Port ReplaceAttentionMaskValue to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
2c8a4e9
Port RMSNorm/LayerNorm surgeries to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
3a0e95a
Port MatMulAddToGemm and GemmToMatMulAdd to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
f961cfa
Port GQA/quant graph surgeries to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
a658b1e
Port QuantizeEmbeddingInt8 and ShareEmbeddingLmHead to onnx_ir

Justin Chu (justinchuby)
and
Copilot
committed
4591b20
Add tests for PowReduceSumPowDiv2LpNorm and QuantizeEmbeddingInt8

Justin Chu (justinchuby)
and
Copilot
committed
85442ef
Address PR review comments on graph surgery migration

Justin Chu (justinchuby)
and
Copilot
committed
838de67
Reuse quantized embedding table for tied LM head in TieWordEmbeddings

Justin Chu (justinchuby)
and
Copilot
committed
3dcb3fa