Quantize and tie q4f16 input embedding

Quantizes the q4f16 input embedding with GatherBlockQuantized and lets its differently shaped initializer share the LM-head packed external byte range. This reduces external data from 2.058 GB to 1.534 GB (decimal) without introducing duplicate tied-weight storage.

Validation:
- shared-range graph is bit-exact to the equivalent duplicate-storage graph for prefill logits, cached logits, and all cache tensors
- ONNX Runtime CPU session creation and Transformers.js WebGPU loading/generation pass
- embedding quantization matched 11/12 varied 32-token greedy generations versus the FP16-embedding graph; the README documents the quantized embedding
- direct WebGPU benchmark improved from 102.16 to 106.05 tokens/s in the local test environment

README.md

@@ -56,14 +56,14 @@ Find more details in the original model card: https://huggingface.co/LiquidAI/LF
 | Precision | Size  | Platform        | Use Case |
 |-----------|-------|-----------------|----------|
 | Q4 | ~1.9 GB | WebGPU, Server | Recommended for most uses (quantized embedding) |
-| Q4F16 | ~2.1 GB | WebGPU | Optimized q4 weights with FP16 runtime and caches |
+| Q4F16 | ~1.5 GB | WebGPU | Quantized embedding and q4 weights with FP16 runtime and caches |
 | FP16 | ~2.1 GB | WebGPU, Server | Higher quality |
 | Q8 | ~2.1 GB | Server only | Balance of quality and size |
 
 - **WebGPU**: Use `Q4`, `Q4F16`, or `FP16` (`Q8` is not supported on WebGPU).
 - **Server (CPU/GPU)**: All variants supported.
 
-Q4 uses a quantized input embedding. Q4F16 keeps the input embedding in FP16 while quantizing the LM head and decoder linear weights to q4.
+Q4 and Q4F16 use a quantized input embedding. Q4F16 uses FP16 runtime tensors and caches while quantizing the LM head and decoder linear weights to q4.
 
 ## Model Files
 
@@ -72,7 +72,7 @@ onnx/
 ├── model.onnx              # FP32
 ├── model_fp16.onnx         # FP16
 ├── model_q4.onnx           # Q4, quantized embedding (WebGPU)
-├── model_q4f16.onnx        # Q4 weights, FP16 runtime and caches (WebGPU)
+├── model_q4f16.onnx        # Q4 embedding/weights, FP16 runtime and caches (WebGPU)
 └── model_q8.onnx           # Q8
 ```
 

onnx/model_q4f16.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:34d1e9d9a5c72659e338c8de3c11302b27d78d009fe891e6c060bec8246767a1
-size 222663
+oid sha256:ab9e9e77e202d7bd2b8a994ac8383f535a4277b4124d5621c79ccc5bb1508bde
+size 222283

onnx/model_q4f16.onnx_data

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:13bd61df0b2449c3bce67c12f1eed81dd27341add0eb9d4caa60e7804b46861b
-size 1072164864
+oid sha256:34537bf4a6d70ddf1627bd3709d31c8b7db8d5bcaee2098c45661be59476fbec
+size 1063972864

onnx/model_q4f16.onnx_data_1

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5ad378ec208f81195cd482efc0d5f7b91c1342f6c73bfea05dbddd7c9302de31
-size 986054656
+oid sha256:1b645b44902caccd406098c4dbef5724927c5fb2a2be4a087ae328989b111a7f
+size 469958656