Split out the quantized file. (#456)

candle-core/src/quantized/mod.rs

@@ -0,0 +1,82 @@
+use crate::Result;
+
+pub mod ggml_file;
+pub mod k_quants;
+
+pub use k_quants::GgmlType;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum GgmlDType {
+    F32,
+    F16,
+    Q4_0,
+    Q4_1,
+    Q5_0,
+    Q5_1,
+    Q8_0,
+    Q8_1,
+    Q2K,
+    Q3K,
+    Q4K,
+    Q5K,
+    Q6K,
+    Q8K,
+}
+
+impl GgmlDType {
+    pub(crate) fn from_u32(u: u32) -> Result<Self> {
+        let dtype = match u {
+            0 => Self::F32,
+            1 => Self::F16,
+            2 => Self::Q4_0,
+            3 => Self::Q4_1,
+            6 => Self::Q5_0,
+            7 => Self::Q5_1,
+            8 => Self::Q8_0,
+            9 => Self::Q8_1,
+            10 => Self::Q2K,
+            11 => Self::Q3K,
+            12 => Self::Q4K,
+            13 => Self::Q5K,
+            14 => Self::Q6K,
+            15 => Self::Q8K,
+            _ => crate::bail!("unknown dtype for tensor {u}"),
+        };
+        Ok(dtype)
+    }
+
+    fn type_size(&self) -> usize {
+        use k_quants::*;
+        match self {
+            Self::F32 => 4,
+            Self::F16 => 2,
+            Self::Q4_0 => std::mem::size_of::<BlockQ4_0>(),
+            Self::Q4_1 => std::mem::size_of::<BlockQ4_1>(),
+            Self::Q5_0 => std::mem::size_of::<BlockQ5_0>(),
+            Self::Q5_1 => std::mem::size_of::<BlockQ5_1>(),
+            // https://github.com/ggerganov/llama.cpp/blob/468ea24fb4633a0d681f7ac84089566c1c6190cb/ggml.c#L932
+            Self::Q8_0 => std::mem::size_of::<BlockQ8_0>(),
+            Self::Q8_1 => std::mem::size_of::<BlockQ8_1>(),
+            Self::Q2K => std::mem::size_of::<BlockQ2K>(),
+            Self::Q3K => std::mem::size_of::<BlockQ3K>(),
+            Self::Q4K => std::mem::size_of::<BlockQ4K>(),
+            Self::Q5K => std::mem::size_of::<BlockQ5K>(),
+            Self::Q6K => std::mem::size_of::<BlockQ6K>(),
+            Self::Q8K => std::mem::size_of::<BlockQ8K>(),
+        }
+    }
+
+    fn blck_size(&self) -> usize {
+        match self {
+            Self::F32 => 1,
+            Self::F16 => 1,
+            Self::Q4_0 => k_quants::QK4_0,
+            Self::Q4_1 => k_quants::QK4_1,
+            Self::Q5_0 => k_quants::QK5_0,
+            Self::Q5_1 => k_quants::QK5_1,
+            Self::Q8_0 => k_quants::QK8_0,
+            Self::Q8_1 => k_quants::QK8_1,
+            Self::Q2K | Self::Q3K | Self::Q4K | Self::Q5K | Self::Q6K | Self::Q8K => k_quants::QK_K,
+        }
+    }
+}