Merge branch 'main' of github.com:AnubhabB/candle into ones-impl

huggingface · Oct 1, 2024 · 942a617 · 942a617
2 parents 453ae03 + def4c6c
commit 942a617
Show file tree

Hide file tree

Showing 32 changed files with 2,712 additions and 205 deletions.
diff --git a/Cargo.toml b/Cargo.toml
@@ -20,7 +20,7 @@ exclude = [
 resolver = "2"
 
 [workspace.package]
-version = "0.7.1"
+version = "0.7.2"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -33,14 +33,14 @@ ab_glyph = "0.2.23"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
-candle = { path = "./candle-core", package = "candle-core", version = "0.7.1" }
-candle-datasets = { path = "./candle-datasets", version = "0.7.1" }
-candle-flash-attn = { path = "./candle-flash-attn", version = "0.7.1" }
-candle-kernels = { path = "./candle-kernels", version = "0.7.1" }
-candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.7.1" }
-candle-nn = { path = "./candle-nn", version = "0.7.1" }
-candle-onnx = { path = "./candle-onnx", version = "0.7.1" }
-candle-transformers = { path = "./candle-transformers", version = "0.7.1" }
+candle = { path = "./candle-core", package = "candle-core", version = "0.7.2" }
+candle-datasets = { path = "./candle-datasets", version = "0.7.2" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.7.2" }
+candle-kernels = { path = "./candle-kernels", version = "0.7.2" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.7.2" }
+candle-nn = { path = "./candle-nn", version = "0.7.2" }
+candle-onnx = { path = "./candle-onnx", version = "0.7.2" }
+candle-transformers = { path = "./candle-transformers", version = "0.7.2" }
 clap = { version = "4.2.4", features = ["derive"] }
 criterion = { version = "0.5.1", default-features=false }
 cudarc = { version = "0.12.1", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }

diff --git a/candle-core/src/quantized/cuda.rs b/candle-core/src/quantized/cuda.rs
@@ -6,9 +6,15 @@ use half::f16;
 
 use cudarc::driver::{CudaSlice, CudaView, DeviceSlice};
 
+#[derive(Clone, Debug)]
+struct PaddedCudaSlice {
+    inner: CudaSlice<u8>,
+    len: usize,
+}
+
 #[derive(Clone, Debug)]
 pub struct QCudaStorage {
-    data: CudaSlice<u8>,
+    data: PaddedCudaSlice,
     dtype: GgmlDType,
     device: CudaDevice,
 }
@@ -61,7 +67,7 @@ fn quantize_q8_1(
 }
 
 fn dequantize_f32(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     dtype: GgmlDType,
     elem_count: usize,
     dev: &CudaDevice,
@@ -104,21 +110,21 @@ fn dequantize_f32(
     };
 
     if is_k {
-        let params = (data, &dst);
+        let params = (&data.inner, &dst);
         unsafe { func.launch(cfg, params) }.w()?;
     } else {
         let nb32 = match dtype {
             GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
             _ => elem_count / 32,
         };
-        let params = (data, &dst, nb32 as i32);
+        let params = (&data.inner, &dst, nb32 as i32);
         unsafe { func.launch(cfg, params) }.w()?;
     }
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn dequantize_f16(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     dtype: GgmlDType,
     elem_count: usize,
     dev: &CudaDevice,
@@ -161,21 +167,21 @@ fn dequantize_f16(
     };
 
     if is_k {
-        let params = (data, &dst);
+        let params = (&data.inner, &dst);
         unsafe { func.launch(cfg, params) }.w()?;
     } else {
         let nb32 = match dtype {
             GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
             _ => elem_count / 32,
         };
-        let params = (data, &dst, nb32 as i32);
+        let params = (&data.inner, &dst, nb32 as i32);
         unsafe { func.launch(cfg, params) }.w()?;
     }
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn dequantize_mul_mat_vec(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     ncols: usize,
@@ -184,7 +190,7 @@ fn dequantize_mul_mat_vec(
 ) -> Result<CudaStorage> {
     use cudarc::driver::LaunchAsync;
 
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < ncols * nrows {
         crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
     }
@@ -213,13 +219,13 @@ fn dequantize_mul_mat_vec(
         shared_mem_bytes: 0,
     };
 
-    let params = (data, y, &dst, ncols as i32, nrows as i32);
+    let params = (&data.inner, y, &dst, ncols as i32, nrows as i32);
     unsafe { func.launch(cfg, params) }.w()?;
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn mul_mat_vec_via_q8_1(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     ncols: usize,
@@ -229,7 +235,7 @@ fn mul_mat_vec_via_q8_1(
 ) -> Result<CudaStorage> {
     use cudarc::driver::LaunchAsync;
 
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < ncols * nrows {
         crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
     }
@@ -276,7 +282,7 @@ fn mul_mat_vec_via_q8_1(
     };
 
     let params = (
-        data,
+        &data.inner,
         &y_q8_1,
         &dst,
         /* ncols_x */ ncols as i32,
@@ -290,7 +296,7 @@ fn mul_mat_vec_via_q8_1(
 
 #[allow(clippy::too_many_arguments)]
 fn mul_mat_via_q8_1(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     x_rows: usize,
@@ -301,7 +307,7 @@ fn mul_mat_via_q8_1(
 ) -> Result<CudaStorage> {
     use cudarc::driver::LaunchAsync;
 
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < x_rows * x_cols {
         crate::bail!("unexpected lhs size {}, {x_rows} {x_cols}", data_elems)
     }
@@ -315,7 +321,7 @@ fn mul_mat_via_q8_1(
     // Start by quantizing y
     let k_padded = pad(k, MATRIX_ROW_PADDING);
     let y_size_in_bytes =
-        k_padded * y_rows * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+        k_padded * y_cols * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
     let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
     quantize_q8_1(y, &mut y_q8_1, k, y_cols, dev)?;
 
@@ -345,7 +351,7 @@ fn mul_mat_via_q8_1(
     };
 
     let params = (
-        /* vx */ data,
+        /* vx */ &data.inner,
         /* vy */ &y_q8_1,
         /* dst */ &dst,
         /* ncols_x */ x_cols as i32,
@@ -361,9 +367,14 @@ fn mul_mat_via_q8_1(
 impl QCudaStorage {
     pub fn zeros(device: &CudaDevice, el_count: usize, dtype: GgmlDType) -> Result<Self> {
         let size_in_bytes = ceil_div(el_count, dtype.block_size()) * dtype.type_size();
-        let data = device.alloc_zeros::<u8>(size_in_bytes).w()?;
+        let padded_size_in_bytes =
+            ceil_div(el_count + MATRIX_ROW_PADDING, dtype.block_size()) * dtype.type_size();
+        let inner = device.alloc_zeros::<u8>(padded_size_in_bytes).w()?;
         Ok(QCudaStorage {
-            data,
+            data: PaddedCudaSlice {
+                inner,
+                len: size_in_bytes,
+            },
             device: device.clone(),
             dtype,
         })
@@ -403,7 +414,10 @@ impl QCudaStorage {
         }
         // Run the dequantization on cpu.
 
-        let buffer = self.device.dtoh_sync_copy(&self.data).w()?;
+        let buffer = self
+            .device
+            .dtoh_sync_copy(&self.data.inner.slice(..self.data.len))
+            .w()?;
         let mut out = vec![0.0; elem_count];
         let block_len = elem_count / self.dtype.block_size();
         match self.dtype {
@@ -444,13 +458,21 @@ impl QCudaStorage {
         let mut qcpu_storage = crate::Device::Cpu.qzeros(src_len, self.dtype)?;
         qcpu_storage.quantize(&src)?;
         let data = qcpu_storage.data()?;
-        let data = self.device.htod_sync_copy(data.as_ref()).w()?;
-        self.data = data;
+        let padded_len =
+            data.len() + MATRIX_ROW_PADDING * self.dtype.type_size() / self.dtype.block_size();
+        let mut inner = unsafe { self.device.alloc::<u8>(padded_len).w()? };
+        self.device
+            .htod_sync_copy_into(data.as_ref(), &mut inner.slice_mut(..data.len()))
+            .w()?;
+        self.data = PaddedCudaSlice {
+            inner,
+            len: data.len(),
+        };
         Ok(())
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
-        self.data.len()
+        self.data.len
     }
 
     pub fn fwd(
@@ -573,11 +595,19 @@ pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
     let data = unsafe {
         std::slice::from_raw_parts(data.as_ptr() as *const u8, core::mem::size_of_val(data))
     };
-    let data = device.htod_sync_copy(data).w()?;
+    let dtype = T::DTYPE;
+    let padded_len = data.len() + MATRIX_ROW_PADDING * dtype.type_size() / dtype.block_size();
+    let mut inner = unsafe { device.alloc::<u8>(padded_len).w()? };
+    device
+        .htod_sync_copy_into(data, &mut inner.slice_mut(..data.len()))
+        .w()?;
     Ok(QStorage::Cuda(QCudaStorage {
-        data,
+        data: PaddedCudaSlice {
+            inner,
+            len: data.len(),
+        },
         device: device.clone(),
-        dtype: T::DTYPE,
+        dtype,
     }))
 }
 
@@ -677,4 +707,28 @@ mod test {
         assert_eq!(vs[15], 13138824.0);
         Ok(())
     }
+
+    // The following test used to fail under compute-sanitizer until #2526.
+    #[test]
+    fn cuda_mm_q8_1_pad() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let (x_rows, ncols, y_cols) = (4, 16, 2048);
+        let vs: Vec<f32> = (0..ncols * y_cols).map(|v| v as f32 / 256.).collect();
+        let y = dev.htod_sync_copy(&vs).w()?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols * x_rows, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* x_rows */ x_rows,
+            /* x_cols */ ncols,
+            /* y_rows */ ncols,
+            /* y_cols */ y_cols,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let _vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+        Ok(())
+    }
 }
diff --git a/candle-examples/Cargo.toml b/candle-examples/Cargo.toml
@@ -36,6 +36,7 @@ serde_json = { workspace = true }
 symphonia = { version = "0.5.3", features = ["all"], optional = true }
 tokenizers = { workspace = true, features = ["onig"] }
 cpal = { version = "0.15.2", optional = true }
+pdf2image = { version = "0.1.2" , optional = true}
 
 [dev-dependencies]
 anyhow = { workspace = true }
@@ -65,7 +66,7 @@ mkl = ["dep:intel-mkl-src", "candle/mkl", "candle-nn/mkl", "candle-transformers/
 nccl = ["cuda", "cudarc/nccl", "dep:half"]
 onnx = ["candle-onnx"]
 metal = ["candle/metal", "candle-nn/metal"]
-microphone = ["cpal"]
+microphone = ["cpal", "rubato"]
 encodec = ["cpal", "symphonia", "rubato"]
 mimi = ["cpal", "symphonia", "rubato"]
 depth_anything_v2 = ["palette", "enterpolation"]
@@ -117,3 +118,7 @@ required-features = ["depth_anything_v2"]
 [[example]]
 name = "silero-vad"
 required-features = ["onnx"]
+
+[[example]]
+name = "colpali"
+required-features = ["pdf2image"]