Add the cuda dequantize f16 kernels. (#2137)

* Add the cuda dequantize f16 kernels.

* Expose the cuda kernels.

* Add some testing + fix.

* Test the other cases too.

* A few more tests.

* Add an environment variable to enable the dequantize f16 + matmul behavior.

3 files changed, 122 insertions, 17 deletions

diff --git a/candle-core/src/quantized/cuda.rs b/candle-core/src/quantized/cuda.rs
index 5481ca3c..8e4884b2 100644
--- a/candle-core/src/quantized/cuda.rs
+++ b/candle-core/src/quantized/cuda.rs
@@ -2,6 +2,7 @@ use super::{GgmlDType, QStorage};
 use crate::quantized::k_quants::GgmlType;
 use crate::{backend::BackendDevice, cuda_backend::WrapErr};
 use crate::{CudaDevice, CudaStorage, Result};
+use half::f16;
 
 use cudarc::driver::{CudaSlice, CudaView, DeviceSlice};
 
@@ -59,7 +60,7 @@ fn quantize_q8_1(
     Ok(())
 }
 
-fn dequantize(
+fn dequantize_f32(
     data: &CudaSlice<u8>,
     dtype: GgmlDType,
     elem_count: usize,
@@ -69,27 +70,27 @@ fn dequantize(
 
     let nb = (elem_count + 255) / 256;
     let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
-        GgmlDType::Q4_0 => ("dequantize_block_q4_0", false, 32, nb),
-        GgmlDType::Q4_1 => ("dequantize_block_q4_1", false, 32, nb),
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f32", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f32", false, 32, nb),
         GgmlDType::Q5_0 => (
-            "dequantize_block_q5_0",
+            "dequantize_block_q5_0_f32",
             false,
             CUDA_DEQUANTIZE_BLOCK_SIZE,
             ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
         ),
         GgmlDType::Q5_1 => (
-            "dequantize_block_q5_1",
+            "dequantize_block_q5_1_f32",
             false,
             CUDA_DEQUANTIZE_BLOCK_SIZE,
             ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
         ),
-        GgmlDType::Q8_0 => ("dequantize_block_q8_0", false, 32, nb),
-        GgmlDType::Q2K => ("dequantize_block_q2_K", true, 64, nb),
-        GgmlDType::Q3K => ("dequantize_block_q3_K", true, 64, nb),
-        GgmlDType::Q4K => ("dequantize_block_q4_K", true, 32, nb),
-        GgmlDType::Q5K => ("dequantize_block_q5_K", true, 64, nb),
-        GgmlDType::Q6K => ("dequantize_block_q6_K", true, 64, nb),
-        GgmlDType::Q8K => ("dequantize_block_q8_K", true, 32, nb),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f32", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f32", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f32", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f32", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f32", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f32", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f32", true, 32, nb),
         _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
     };
     let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
@@ -116,6 +117,63 @@ fn dequantize(
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
+fn dequantize_f16(
+    data: &CudaSlice<u8>,
+    dtype: GgmlDType,
+    elem_count: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let nb = (elem_count + 255) / 256;
+    let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f16", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f16", false, 32, nb),
+        GgmlDType::Q5_0 => (
+            "dequantize_block_q5_0_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q5_1 => (
+            "dequantize_block_q5_1_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f16", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f16", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f16", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f16", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f16", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f16", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f16", true, 32, nb),
+        _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f16>(elem_count).w()? };
+    // See e.g.
+    // https://github.com/ggerganov/llama.cpp/blob/cbbd1efa06f8c09f9dff58ff9d9af509cc4c152b/ggml-cuda.cu#L7270
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (num_blocks as u32, 1, 1),
+        block_dim: (block_dim as u32, 1, 1),
+        shared_mem_bytes: 0,
+    };
+
+    if is_k {
+        let params = (data, &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);
+        unsafe { func.launch(cfg, params) }.w()?;
+    }
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
 fn dequantize_mul_mat_vec(
     data: &CudaSlice<u8>,
     y: &CudaView<f32>,
@@ -341,7 +399,7 @@ impl QCudaStorage {
                 | GgmlDType::Q8K
         );
         if fast_kernel {
-            return dequantize(&self.data, self.dtype, elem_count, self.device());
+            return dequantize_f32(&self.data, self.dtype, elem_count, self.device());
         }
         // Run the dequantization on cpu.
 
@@ -369,6 +427,10 @@ impl QCudaStorage {
             .storage_from_cpu_storage(&crate::CpuStorage::F32(out))
     }
 
+    pub fn dequantize_f16(&self, elem_count: usize) -> Result<CudaStorage> {
+        dequantize_f16(&self.data, self.dtype, elem_count, self.device())
+    }
+
     pub fn quantize(&mut self, src: &CudaStorage) -> Result<()> {
         // Run the quantization on cpu.
         let src = match &src.slice {
diff --git a/candle-core/src/quantized/dummy_cuda.rs b/candle-core/src/quantized/dummy_cuda.rs
index 598c5cd1..ca7b8120 100644
--- a/candle-core/src/quantized/dummy_cuda.rs
+++ b/candle-core/src/quantized/dummy_cuda.rs
@@ -24,6 +24,10 @@ impl QCudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    pub fn dequantize_f16(&self, _elem_count: usize) -> Result<CudaStorage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     pub fn quantize(&mut self, _src: &CudaStorage) -> Result<()> {
         Err(Error::NotCompiledWithCudaSupport)
     }
diff --git a/candle-core/src/quantized/mod.rs b/candle-core/src/quantized/mod.rs
index 47307f2e..e87072bb 100644
--- a/candle-core/src/quantized/mod.rs
+++ b/candle-core/src/quantized/mod.rs
@@ -1,4 +1,4 @@
-use crate::{CpuStorage, Device, Result, Shape, Storage, Tensor};
+use crate::{CpuStorage, DType, Device, Result, Shape, Storage, Tensor};
 use k_quants::*;
 use std::borrow::Cow;
 
@@ -360,9 +360,24 @@ impl QTensor {
     pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
         let storage = self.storage.dequantize(self.shape.elem_count())?;
         let none = crate::op::BackpropOp::none();
-        let is_variable = false;
-        crate::tensor::from_storage(storage, self.shape.clone(), none, is_variable)
-            .to_device(device)
+        crate::tensor::from_storage(storage, self.shape.clone(), none, false).to_device(device)
+    }
+
+    pub fn dequantize_f16(&self, device: &Device) -> Result<Tensor> {
+        // In the CUDA case, we have a specialized kernel as this can be useful for volta
+        // architectures. https://github.com/huggingface/candle/issues/2136
+        match &self.storage {
+            QStorage::Cuda(s) => {
+                let s = s.dequantize_f16(self.shape.elem_count())?;
+                let none = crate::op::BackpropOp::none();
+                crate::tensor::from_storage(Storage::Cuda(s), self.shape.clone(), none, false)
+                    .to_device(device)
+            }
+            _ => {
+                let s = self.dequantize(device)?.to_dtype(crate::DType::F16)?;
+                Ok(s)
+            }
+        }
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
@@ -378,6 +393,7 @@ impl QTensor {
 pub enum QMatMul {
     QTensor(std::sync::Arc<QTensor>),
     Tensor(Tensor),
+    TensorF16(Tensor),
 }
 
 thread_local! {
@@ -391,6 +407,17 @@ thread_local! {
     }
 }
 
+thread_local! {
+    static DEQUANTIZE_ALL_F16: bool = {
+        match std::env::var("CANDLE_DEQUANTIZE_ALL_F16") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
+
 impl QMatMul {
     pub fn from_arc(qtensor: std::sync::Arc<QTensor>) -> Result<Self> {
         let dequantize = match qtensor.dtype() {
@@ -400,6 +427,9 @@ impl QMatMul {
         let t = if dequantize {
             let tensor = qtensor.dequantize(&qtensor.device())?;
             Self::Tensor(tensor)
+        } else if DEQUANTIZE_ALL_F16.with(|b| *b) {
+            let tensor = qtensor.dequantize_f16(&qtensor.device())?;
+            Self::TensorF16(tensor)
         } else {
             Self::QTensor(qtensor)
         };
@@ -486,6 +516,15 @@ impl crate::Module for QMatMul {
                 };
                 xs.matmul(&w)
             }
+            Self::TensorF16(w) => {
+                let in_dtype = xs.dtype();
+                let w = match *xs.dims() {
+                    [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+                    [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+                    _ => w.t()?,
+                };
+                xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
+            }
         }
     }
 }