Remove the Reshape that duplicated the embedding at load

README.md

@@ -26,6 +26,7 @@ This repository contains modified files. Compared with [onnx-community/LFM3-1.7B
 
 - **Re-quantized weights:** all `MatMulNBits` weights were re-quantized from the fp16 ONNX model with GPTQ (4-bit, block size 128, with zero points). The tied embedding / LM head uses round-to-nearest 4-bit, block size 128, with a clip search.
 - **Shorter rotary caches:** the cos/sin caches were cut from 262,144 to 16,384 positions. Sequences longer than 16,384 tokens fail with a clear error instead of running.
+- **No second copy of the embedding at load:** in the original graph, a `Reshape` feeds the shared embedding / LM-head weights to the LM head, which makes onnxruntime-web copy that 131 MB tensor while creating the session (about +390 MB of wasm heap and +131 MB of GPU memory). The graph now gives the LM head the same weights without the `Reshape`. Outputs are identical.
 - **Calibration data:** about 36k tokens from the WikiText-2 training split plus chat replies generated by the fp16 model. No calibration text is included in the files.
 
 `config.json`, the tokenizer files and the chat template are unchanged.

onnx/model_q4f16.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:43c35780930fa7a85ab7e06d2078259517c81805eac5f47b7ceb3d3edf89da50
-size 310898
+oid sha256:f4bc2c0f1a3ebf6f32e8da253c1cf01d05b1b8f8100e5ad5ca647de80335882a
+size 310650