From 47d72bd462cbf62d762e089cb83ed8c19f60d535 Mon Sep 17 00:00:00 2001
From: cryscan <cryscan@umich.edu>
Date: Sun, 22 Sep 2024 15:23:59 +0800
Subject: [PATCH] Do not load LoRA for embed.

---
 Cargo.toml            |  3 +--
 src/model/loader.rs   | 25 +++----------------------
 src/runtime/loader.rs | 24 +++---------------------
 3 files changed, 7 insertions(+), 45 deletions(-)

diff --git a/Cargo.toml b/Cargo.toml
index 28117ab..c4b88e4 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -38,7 +38,6 @@ trait-variant = "0.1"
 uid = "0.1"
 wasm-bindgen = "0.2"
 wgpu = "22.1.0"
-pollster = "0.3"
 
 [dependencies.web-rwkv-derive]
 path = "crates/web-rwkv-derive"
@@ -47,7 +46,7 @@ version = "0.2.5"
 [dependencies.tokio]
 default-features = false
 features = ["macros", "rt", "sync", "time"]
-version = "1.37"
+version = "1.40"
 
 [dev-dependencies]
 cbor4ii = { version = "0.3.2", features = ["half-f16", "serde1"] }
diff --git a/src/model/loader.rs b/src/model/loader.rs
index e61b31b..ab4eb1a 100644
--- a/src/model/loader.rs
+++ b/src/model/loader.rs
@@ -608,28 +608,9 @@ impl<R: Reader> Loader<R> {
     }
 
     pub fn load_embed(&self) -> Result<TensorCpu<f16>> {
-        let context = &self.context;
-        let name = "emb.weight";
-
-        let (dt, shape, tensor) = self.model.tensor(name)?;
-        let lora = self.lora_vectors(name)?;
-
-        if lora.is_empty() {
-            let tensor = TensorCpu::from_reader((dt, shape, tensor))?;
-            Ok(tensor)
-        } else {
-            let tensor = TensorCpu::from_reader((dt, shape, tensor))?.transfer_into(context);
-            let mut ops = vec![];
-            for lora in lora {
-                let factor = vec![lora.alpha, 1.0, 0.0, 0.0];
-                let factor = context.tensor_from_data([4, 1, 1, 1], factor)?;
-                let op = TensorOp::blend(&factor, &lora.tensor, &tensor)?;
-                ops.push(op);
-            }
-
-            context.queue.submit(context.encode(&TensorOp::List(ops)));
-            Ok(pollster::block_on(tensor.back()))
-        }
+        let (dt, shape, tensor) = self.model.tensor("emb.weight")?;
+        let tensor = TensorCpu::from_reader((dt, shape, tensor))?;
+        Ok(tensor)
     }
 
     pub fn load_head(&self, chunk_size: usize) -> Result<Vec<TensorGpu<f16, ReadWrite>>> {
diff --git a/src/runtime/loader.rs b/src/runtime/loader.rs
index 16235ca..b7d0a97 100644
--- a/src/runtime/loader.rs
+++ b/src/runtime/loader.rs
@@ -607,27 +607,9 @@ impl<R: Reader> Loader<R> {
     }
 
     pub fn load_embed(&self) -> Result<TensorCpu<f16>> {
-        let context = &self.context;
-        let name = "emb.weight";
-
-        let (dt, shape, tensor) = self.model.tensor(name)?;
-        let lora = self.lora_vectors(name)?;
-
-        if lora.is_empty() {
-            let tensor = TensorCpu::from_reader((dt, shape, tensor))?;
-            Ok(tensor)
-        } else {
-            let tensor = TensorCpu::from_reader((dt, shape, tensor))?.transfer_into(context);
-            let mut ops = vec![];
-            for lora in lora {
-                let factor = vec![lora.alpha, 1.0, 0.0, 0.0];
-                let factor = context.tensor_from_data([4, 1, 1, 1], factor)?;
-                let op = TensorOp::blend(&factor, &lora.tensor, &tensor)?;
-                ops.push(op);
-            }
-            context.queue.submit(context.encode(&TensorOp::List(ops)));
-            Ok(pollster::block_on(tensor.back()))
-        }
+        let (dt, shape, tensor) = self.model.tensor("emb.weight")?;
+        let tensor = TensorCpu::from_reader((dt, shape, tensor))?;
+        Ok(tensor)
     }
 
     pub fn load_head(&self, chunk_size: usize) -> Result<Vec<TensorGpu<f16, ReadWrite>>> {