rublas 0.2.8

Linear algebra and matrix multiplication for Ruda.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242

mod codebooks;
mod kernels;
#[cfg(test)]
mod tests;

use half::{bf16, f16};
use ruda_core::{
    device::Device,
    tensor::{DType, Shape, data::TensorData, quantization::GgmlType},
};
use ruda_kernel::{
    dsl::{Runtime, calculate_ruda_count_elemwise, prelude::{ArrayArg, RudaDim}},
    tensor::{RudaTensor, allocation::empty_device_contiguous_dtype, contiguous::into_contiguous, transfer::from_data},
};

#[derive(Clone, Debug, PartialEq, Eq)]
pub struct GgmlError(pub String);
impl core::fmt::Display for GgmlError {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { f.write_str(&self.0) }
}
impl std::error::Error for GgmlError {}

/// Original GGML block storage in logical `[rows, width]` order.
/// Only the packed file bytes and the format's small shared codebook are resident.
#[derive(Clone, Debug)]
pub struct GgmlGemm<R: Runtime> {
    packed: RudaTensor<R>,
    table: RudaTensor<R>,
    format: GgmlType,
    big_endian: bool,
    rows: usize,
    width: usize,
    row_bytes: usize,
}

impl<R: Runtime> GgmlGemm<R> {
    pub fn from_bytes(
        data: Vec<u8>, format: GgmlType, rows: usize, width: usize,
        big_endian: bool, device: &R::Device,
    ) -> Result<Self, GgmlError> {
        let row_bytes = format.row_bytes(width).ok_or_else(|| GgmlError("GGML row is not block aligned".into()))?;
        if rows == 0 || width == 0 || rows.checked_mul(row_bytes) != Some(data.len())
            || data.len() > u32::MAX as usize || width > u32::MAX as usize || rows > u32::MAX as usize
        { return Err(GgmlError("GGML packed geometry or byte length is invalid".into())); }
        if !format.is_quantized() && !matches!(format, GgmlType::F32 | GgmlType::F16 | GgmlType::BF16) {
            return Err(GgmlError(format!("{format:?} is not a supported floating GGML projection format")));
        }
        let table: Vec<u8> = match format {
            GgmlType::IQ2_XXS => codebooks::IQ2XXS_GRID.iter().flat_map(|n| n.to_le_bytes()).collect(),
            GgmlType::IQ2_XS => codebooks::IQ2XS_GRID.iter().flat_map(|n| n.to_le_bytes()).collect(),
            GgmlType::IQ2_S => codebooks::IQ2S_GRID.iter().flat_map(|n| n.to_le_bytes()).collect(),
            GgmlType::IQ3_XXS => codebooks::IQ3XXS_GRID.iter().flat_map(|n| n.to_le_bytes()).collect(),
            GgmlType::IQ3_S => codebooks::IQ3S_GRID.iter().flat_map(|n| n.to_le_bytes()).collect(),
            GgmlType::IQ1_S | GgmlType::IQ1_M => codebooks::IQ1S_GRID.iter().flat_map(|n| n.to_le_bytes()).collect(),
            GgmlType::IQ4_NL | GgmlType::IQ4_XS => codebooks::KVALUES_IQ4NL.iter().map(|n| *n as u8).collect(),
            GgmlType::MXFP4 | GgmlType::NVFP4 => codebooks::KVALUES_FP4.iter().map(|n| *n as u8).collect(),
            _ => vec![0u8],
        };
        let size = data.len();
        let packed = from_data(TensorData::new(data, [size]), device);
        let table_len = table.len();
        let table = from_data(TensorData::new(table, [table_len]), device);
        Ok(Self { packed, table, format, big_endian, rows, width, row_bytes })
    }

    pub fn format(&self) -> GgmlType { self.format }
    pub fn rows(&self) -> usize { self.rows }
    pub fn width(&self) -> usize { self.width }
    pub fn packed_bytes(&self) -> usize { self.rows * self.row_bytes }
    pub fn original_storage(&self) -> &RudaTensor<R> { &self.packed }

    pub fn apply_vector(&self, input: RudaTensor<R>, multiply: bool) -> Result<RudaTensor<R>, GgmlError> {
        self.same_device(&input)?;
        let shape = input.meta.shape().clone();
        let rank = shape.num_dims();
        if self.rows != 1 || rank == 0 || shape[rank - 1] != self.width
            || !matches!(input.dtype, DType::F32 | DType::F16 | DType::BF16)
            || input.meta.num_elements() > u32::MAX as usize
        { return Err(GgmlError("GGML vector operation requires matching floating input".into())); }
        let input = into_contiguous(input);
        let size = input.meta.num_elements();
        let output = empty_device_contiguous_dtype(input.client.clone(), input.device.clone(), shape, input.dtype);
        if size != 0 {
            let dim = RudaDim::new(input.client.properties(), size);
            let grid = calculate_ruda_count_elemwise(&input.client, size, dim);
            let (block, bytes) = self.format.block_layout();
            macro_rules! run {
                ($f:ty) => { unsafe { kernels::vector::launch::<$f, R>(
                    &input.client, grid, dim, input.clone().into_tensor_arg(),
                    self.packed.clone().into_tensor_arg(), self.table.clone().into_tensor_arg(),
                    output.clone().into_tensor_arg(), self.width, multiply,
                    block, bytes, self.format as u32, self.big_endian,
                ) } };
            }
            match input.dtype { DType::F32 => run!(f32), DType::F16 => run!(f16), DType::BF16 => run!(bf16), _ => unreachable!() }
        }
        Ok(output)
    }

    fn same_device(&self, tensor: &RudaTensor<R>) -> Result<(), GgmlError> {
        if tensor.device.to_id() != self.packed.device.to_id()
            || !tensor.client.same_execution_queue(&self.packed.client) || tensor.qparams.is_some()
        { return Err(GgmlError("GGML input and packed weight must share device and execution queue".into())); }
        Ok(())
    }

    pub fn forward(&self, input: RudaTensor<R>) -> Result<RudaTensor<R>, GgmlError> {
        self.same_device(&input)?;
        let mut shape = input.meta.shape().clone();
        let rank = shape.num_dims();
        if rank == 0 || shape[rank - 1] != self.width
            || !matches!(input.dtype, DType::F32 | DType::F16 | DType::BF16)
        { return Err(GgmlError("GGML projection requires matching floating input width".into())); }
        shape[rank - 1] = self.rows;
        let size = shape.iter().try_fold(1usize, |n, &d| n.checked_mul(d))
            .ok_or_else(|| GgmlError("GGML output size overflows".into()))?;
        if size > u32::MAX as usize || input.meta.num_elements() > u32::MAX as usize {
            return Err(GgmlError("GGML activation exceeds native indexing range".into()));
        }
        let input = into_contiguous(input);
        let output = empty_device_contiguous_dtype(input.client.clone(), input.device.clone(), shape, input.dtype);
        if size != 0 {
            let dim = RudaDim::new(input.client.properties(), size);
            let grid = calculate_ruda_count_elemwise(&input.client, size, dim);
            let (block, bytes) = self.format.block_layout();
            macro_rules! run {
                ($f:ty) => { unsafe { kernels::matmul::launch::<$f, R>(
                    &input.client, grid, dim, input.clone().into_tensor_arg(),
                    self.packed.clone().into_tensor_arg(), self.table.clone().into_tensor_arg(),
                    output.clone().into_tensor_arg(), self.width, self.rows, self.row_bytes,
                    block, bytes, self.format as u32, self.big_endian,
                ) } };
            }
            match input.dtype { DType::F32 => run!(f32), DType::F16 => run!(f16), DType::BF16 => run!(bf16), _ => unreachable!() }
        }
        Ok(output)
    }

    /// Applies selected expert matrices to `[tokens, 1 | selected, width]`.
    /// Packed rows are `[experts, columns]`; indices are `[tokens, selected]`.
    pub fn forward_selected(
        &self, input: RudaTensor<R>, indices: RudaTensor<R>, experts: usize,
    ) -> Result<RudaTensor<R>, GgmlError> {
        self.same_device(&input)?;
        self.same_device(&indices)?;
        let shape = input.meta.shape();
        let index_shape = indices.meta.shape();
        if experts == 0 || self.rows % experts != 0 || shape.num_dims() != 3
            || index_shape.num_dims() != 2 || shape[0] != index_shape[0]
            || index_shape[1] == 0 || (shape[1] != 1 && shape[1] != index_shape[1])
            || shape[2] != self.width
            || !matches!(input.dtype, DType::F32 | DType::F16 | DType::BF16)
            || !matches!(indices.dtype, DType::I32 | DType::I64)
        { return Err(GgmlError("GGML selected expert geometry or dtype is invalid".into())); }
        let columns = self.rows / experts;
        let selected = index_shape[1];
        let input_slots = shape[1];
        let output_shape = Shape::from([shape[0], selected, columns]);
        let size = shape[0].checked_mul(selected).and_then(|n| n.checked_mul(columns))
            .filter(|n| *n <= u32::MAX as usize)
            .ok_or_else(|| GgmlError("GGML selected expert output exceeds native indexing range".into()))?;
        if input.meta.num_elements() > u32::MAX as usize || indices.meta.num_elements() > u32::MAX as usize {
            return Err(GgmlError("GGML selected expert input exceeds native indexing range".into()));
        }
        let input = into_contiguous(input);
        let indices = into_contiguous(indices);
        let output = empty_device_contiguous_dtype(
            input.client.clone(), input.device.clone(), output_shape, input.dtype);
        if size != 0 {
            let status = input.client.create_from_slice(&0u32.to_ne_bytes());
            let dim = RudaDim::new(input.client.properties(), size);
            let grid = calculate_ruda_count_elemwise(&input.client, size, dim);
            let (block, bytes) = self.format.block_layout();
            macro_rules! run {
                ($f:ty, $i:ty) => { unsafe { kernels::matmul_selected::launch::<$f, $i, R>(
                    &input.client, grid, dim, input.clone().into_tensor_arg(), indices.clone().into_tensor_arg(),
                    self.packed.clone().into_tensor_arg(), self.table.clone().into_tensor_arg(),
                    output.clone().into_tensor_arg(), ArrayArg::from_raw_parts(status.clone(), 1),
                    self.width, columns, experts,
                    selected, input_slots, self.row_bytes, block, bytes, self.format as u32, self.big_endian,
                ) } };
            }
            match (input.dtype, indices.dtype) {
                (DType::F32, DType::I32) => run!(f32, i32), (DType::F32, DType::I64) => run!(f32, i64),
                (DType::F16, DType::I32) => run!(f16, i32), (DType::F16, DType::I64) => run!(f16, i64),
                (DType::BF16, DType::I32) => run!(bf16, i32), (DType::BF16, DType::I64) => run!(bf16, i64),
                _ => unreachable!(),
            }
            let status = input.client.read_one(status)
                .map_err(|error| GgmlError(format!("GGUF index check readback failed: {error}")))?;
            let status: [u8; 4] = status[..].try_into()
                .map_err(|_| GgmlError("GGUF index check returned an invalid status size".into()))?;
            if u32::from_ne_bytes(status) != 0 {
                return Err(GgmlError("GGUF expert index out of bounds".into()));
            }
        }
        Ok(output)
    }

    /// Gathers only requested embeddings; the vocabulary table stays packed.
    pub fn gather(&self, indices: RudaTensor<R>, dtype: DType) -> Result<RudaTensor<R>, GgmlError> {
        self.same_device(&indices)?;
        if !matches!(indices.dtype, DType::I32 | DType::I64)
            || !matches!(dtype, DType::F32 | DType::F16 | DType::BF16)
        { return Err(GgmlError("GGML gather requires integer indices and floating output".into())); }
        let mut shape = indices.meta.shape().to_vec();
        shape.push(self.width);
        let size = indices.meta.num_elements().checked_mul(self.width)
            .filter(|n| *n <= u32::MAX as usize).ok_or_else(|| GgmlError("GGML gather size overflows".into()))?;
        let indices = into_contiguous(indices);
        let output = empty_device_contiguous_dtype(indices.client.clone(), indices.device.clone(), Shape::from(shape), dtype);
        if size != 0 {
            let status = indices.client.create_from_slice(&0u32.to_ne_bytes());
            let dim = RudaDim::new(indices.client.properties(), size);
            let grid = calculate_ruda_count_elemwise(&indices.client, size, dim);
            let (block, bytes) = self.format.block_layout();
            macro_rules! run {
                ($f:ty, $i:ty) => { unsafe { kernels::gather::launch::<$f, $i, R>(
                    &indices.client, grid, dim, indices.clone().into_tensor_arg(),
                    self.packed.clone().into_tensor_arg(), self.table.clone().into_tensor_arg(),
                    output.clone().into_tensor_arg(), ArrayArg::from_raw_parts(status.clone(), 1),
                    self.width, self.rows, self.row_bytes,
                    block, bytes, self.format as u32, self.big_endian,
                ) } };
            }
            match (dtype, indices.dtype) {
                (DType::F32, DType::I32) => run!(f32, i32), (DType::F32, DType::I64) => run!(f32, i64),
                (DType::F16, DType::I32) => run!(f16, i32), (DType::F16, DType::I64) => run!(f16, i64),
                (DType::BF16, DType::I32) => run!(bf16, i32), (DType::BF16, DType::I64) => run!(bf16, i64),
                _ => unreachable!(),
            }
            let status = indices.client.read_one(status)
                .map_err(|error| GgmlError(format!("GGUF index check readback failed: {error}")))?;
            let status: [u8; 4] = status[..].try_into()
                .map_err(|_| GgmlError("GGUF index check returned an invalid status size".into()))?;
            if u32::from_ne_bytes(status) != 0 {
                return Err(GgmlError("GGUF embedding index out of bounds".into()));
            }
        }
        Ok(output)
    }
}