rublas 0.1.110

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

use ruda_core::tensor::{DType, Shape};
use ruda_kernel::{dsl as kernel_dsl, dsl::prelude::*, tensor::{RudaTensor,
    allocation::empty_device_contiguous_dtype, permutation::swap_dims}};
use crate::kernel_ir::{definition::MatmulSetupError, launch::Strategy};
use super::{F32MathMode, launch_matmul_with_precision};

/// A100 Qwen K/V projection: first sliced accumulator plus bias, then second.
pub fn matmul_bf16_sliced_k2_bias<R: Runtime>(a: RudaTensor<R>, b: RudaTensor<R>,
    bias: RudaTensor<R>, out: RudaTensor<R>) -> Result<RudaTensor<R>, MatmulSetupError> {
    if a.dtype != DType::BF16 || b.dtype != DType::BF16 || bias.dtype != DType::BF16 || out.dtype != DType::BF16
        || a.qparams.is_some() || b.qparams.is_some() || bias.qparams.is_some() || out.qparams.is_some()
        || a.meta.num_dims()!=2 || b.meta.num_dims()!=2 || bias.meta.num_dims()!=2 || out.meta.num_dims()!=2
        || a.meta.shape()[..]!=[512,3584] || b.meta.shape()[..]!=[3584,512]
        || bias.meta.shape()[..]!=[512,512]
        || out.meta.shape()[..]!=[512,512] || out.meta.strides()[..]!=[512,1]
        || !a.client.same_execution_queue(&b.client) || !a.client.same_execution_queue(&bias.client)
        || !a.client.same_execution_queue(&out.client)
    {
        return Err(MatmulSetupError::InvalidConfig(Box::new("unsupported BF16 sliced-K2 projection".to_string())));
    }
    let left = empty_device_contiguous_dtype(a.client.clone(),a.device.clone(),Shape::from([2,512,1792]),DType::BF16);
    let right = empty_device_contiguous_dtype(a.client.clone(),a.device.clone(),Shape::from([2,512,1792]),DType::BF16);
    for (input,output,rhs) in [(a.clone(),left.clone(),false),(b,right.clone(),true)] {
        pack::launch::<R>(&a.client,RudaCount::Static((2*512*1792/128) as u32,1,1),RudaDim::new_1d(128),
            input.into_tensor_arg(),output.into_array_arg(),rhs,DType::BF16.into());
    }
    let partial = empty_device_contiguous_dtype(a.client.clone(),a.device.clone(),Shape::from([2,512,512]),DType::F32);
    let right = swap_dims(right, 1, 2);
    if !super::matmul_bf16_training(&left, &right, &partial) {
        launch_matmul_with_precision(&Strategy::DoubleAsyncCyclicCmma(Default::default()),
            left, right, partial.clone(), F32MathMode::Strict)?;
    }
    reduce::launch::<R>(&a.client,RudaCount::Static((512*512/128) as u32,1,1),RudaDim::new_1d(128),
        partial.into_array_arg(),bias.into_tensor_arg(),out.clone().into_array_arg(),DType::BF16.into());
    Ok(out)
}

/// Ordered three-part split-K with BF16 writeback between partial sums.
pub fn matmul_bf16_serial_k3<R: Runtime>(a: RudaTensor<R>, b: RudaTensor<R>,
    out: RudaTensor<R>) -> Result<RudaTensor<R>, MatmulSetupError> {
    split_k(a, b, out, 3, false)
}

/// Round each split-K partial to BF16 before its ordered FP32 reduction.
pub fn matmul_bf16_rounded_parallel_k<R: Runtime>(a: RudaTensor<R>, b: RudaTensor<R>,
    out: RudaTensor<R>, parts: usize) -> Result<RudaTensor<R>, MatmulSetupError> {
    split_k(a, b, out, parts, true)
}

fn split_k<R: Runtime>(a: RudaTensor<R>, b: RudaTensor<R>, out: RudaTensor<R>,
    parts: usize, rounded_partials: bool) -> Result<RudaTensor<R>, MatmulSetupError> {
    if a.dtype != DType::BF16 || b.dtype != DType::BF16 || out.dtype != DType::BF16
        || a.qparams.is_some() || b.qparams.is_some() || out.qparams.is_some()
        || a.meta.num_dims()!=2 || b.meta.num_dims()!=2 || out.meta.num_dims()!=2
        || !matches!(parts, 3 | 5) || (!rounded_partials && parts != 3)
        || b.meta.shape()[0]!=a.meta.shape()[1]
        || out.meta.shape()[..]!=[a.meta.shape()[0],b.meta.shape()[1]]
        || out.meta.strides()[..]!=[b.meta.shape()[1],1]
        || a.meta.strides()[1]!=1 || (b.meta.strides()[0]!=1 && b.meta.strides()[1]!=1)
        || !a.client.same_execution_queue(&b.client) || !a.client.same_execution_queue(&out.client)
    {
        return Err(MatmulSetupError::InvalidConfig(Box::new("unsupported BF16 split-K projection".to_string())));
    }
    let (m, k, n) = (a.meta.shape()[0], a.meta.shape()[1], b.meta.shape()[1]);
    let block = k.div_ceil(parts*64)*64;
    if m == 0 || n == 0 || k == 0 || k % 64 != 0 || k <= (parts-1)*block
        || m.checked_mul(n).and_then(|v| v.checked_mul(parts)).is_none_or(|v| v > u32::MAX as usize)
    {
        return Err(MatmulSetupError::InvalidConfig(Box::new("invalid BF16 split-K extent".to_string())));
    }
    let partial = empty_device_contiguous_dtype(a.client.clone(),a.device.clone(),Shape::from([parts,m,n]),
        if rounded_partials { DType::BF16 } else { DType::F32 });
    let element_bytes = if rounded_partials { 2 } else { 4 };
    if !super::training_gemm::split_k(&a, &b, &partial, parts, block) {
        for part in 0..parts {
            let start = part*block;
            let len = (k-start).min(block);
            let mut left = a.clone();
            left.meta.shape_mut()[1] = len;
            left.handle.offset_start = Some(a.handle.offset_start.unwrap_or(0)+(start*a.meta.strides()[1]*2) as u64);
            let mut right = b.clone();
            right.meta.shape_mut()[0] = len;
            right.handle.offset_start = Some(b.handle.offset_start.unwrap_or(0)+(start*b.meta.strides()[0]*2) as u64);
            let mut output = partial.clone();
            output.meta.remove(0);
            output.handle.offset_start = Some(partial.handle.offset_start.unwrap_or(0)+(part*m*n*element_bytes) as u64);
            if super::matmul_bf16_training(&left, &right, &output) { continue; }
            match launch_matmul_with_precision(&Strategy::DoubleAsyncCyclicCmma(Default::default()),
                left.clone(),right.clone(),output.clone(),F32MathMode::Strict) {
                Ok(()) => {},
                Err(MatmulSetupError::Unavailable(_) | MatmulSetupError::InvalidConfig(_) | MatmulSetupError::Vectorization(_)) => {
                    launch_matmul_with_precision(&Strategy::Auto,left,right,output,F32MathMode::Strict)?;
                },
                Err(error) => return Err(error),
            }
        }
    }
    if rounded_partials {
        rounded_parallel_reduce::launch::<R>(&a.client,RudaCount::Static((m*n).div_ceil(128) as u32,1,1),RudaDim::new_1d(128),
            partial.into_array_arg(),out.clone().into_array_arg(),parts,DType::BF16.into());
    } else {
        serial_reduce::launch::<R>(&a.client,RudaCount::Static((m*n).div_ceil(128) as u32,1,1),RudaDim::new_1d(128),
            partial.into_array_arg(),out.clone().into_array_arg(),DType::BF16.into());
    }
    Ok(out)
}

/// Parallel LM-head input gradients without copying the frozen vocabulary weight.
pub fn matmul_bf16_head_parallel_k6<R: Runtime>(a: RudaTensor<R>, b: RudaTensor<R>,
    out: RudaTensor<R>) -> Result<RudaTensor<R>, MatmulSetupError> {
    if a.dtype!=DType::BF16 || b.dtype!=DType::BF16 || !matches!(out.dtype,DType::BF16|DType::F32)
        || a.qparams.is_some() || b.qparams.is_some() || out.qparams.is_some()
        || a.meta.num_dims()!=2 || b.meta.num_dims()!=2 || out.meta.num_dims()!=2
        || !matches!(a.meta.shape()[0],31|32) || a.meta.shape()[1]!=152064
        || b.meta.shape()[..]!=[152064,3584] || out.meta.shape()[..]!=[a.meta.shape()[0],3584]
        || a.meta.strides()[..]!=[152064,1] || b.meta.strides()[..]!=[3584,1]
        || out.meta.strides()[..]!=[3584,1]
        || a.handle.offset_start.unwrap_or(0)%16!=0 || b.handle.offset_start.unwrap_or(0)%16!=0
        || !a.client.same_execution_queue(&b.client) || !a.client.same_execution_queue(&out.client)
    {
        return Err(MatmulSetupError::InvalidConfig(Box::new("unsupported BF16 parallel LM-head gradient".to_string())));
    }
    let m = a.meta.shape()[0];
    let mut left = a.clone();
    left.meta.shape_mut()[1] = 25344;
    left.meta.insert(0,6,25344);
    let mut right = b;
    right.meta.shape_mut()[0] = 25344;
    right.meta.insert(0,6,25344*3584);
    let partial = empty_device_contiguous_dtype(a.client.clone(),a.device.clone(),Shape::from([6,m,3584]),DType::F32);
    if !super::matmul_bf16_training(&left, &right, &partial) {
        launch_matmul_with_precision(&Strategy::Auto,left,right,partial.clone(),F32MathMode::Strict)?;
    }
    parallel_reduce::launch::<R>(&a.client,RudaCount::Static((m*3584).div_ceil(128) as u32,1,1),RudaDim::new_1d(128),
        partial.into_array_arg(),out.clone().into_array_arg(),out.dtype.into());
    Ok(out)
}

#[ruda(launch)]
fn parallel_reduce<F: Float>(partial: &Array<f32>, out: &mut Array<F>, #[define(F)] _dtype: StorageType) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() {
        let mut value = partial[pos];
        #[unroll]
        for part in 1usize..6 { value += partial[part*out.len()+pos]; }
        out[pos] = F::cast_from(value);
    }
}

#[ruda(launch)]
fn rounded_parallel_reduce<F: Float>(partial: &Array<F>, out: &mut Array<F>,
    #[comptime] parts: usize, #[define(F)] _dtype: StorageType) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() {
        let mut value = f32::cast_from(partial[pos]);
        #[unroll]
        for part in 1usize..parts { value += f32::cast_from(partial[part*out.len()+pos]); }
        out[pos] = F::cast_from(value);
    }
}

#[ruda(launch)]
fn serial_reduce<F: Float>(partial: &Array<f32>, out: &mut Array<F>, #[define(F)] _dtype: StorageType) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() {
        let first = F::cast_from(partial[pos]);
        let second = F::cast_from(f32::cast_from(first)+partial[out.len()+pos]);
        out[pos] = F::cast_from(f32::cast_from(second)+partial[2*out.len()+pos]);
    }
}

#[ruda(launch)]
fn pack<F: Float>(input: &Tensor<F>, output: &mut Array<F>,
    #[comptime] rhs: bool, #[define(F)] _dtype: StorageType) {
    let pos = ABSOLUTE_POS as usize;
    if pos < output.len() {
        let part = pos / (512*1792);
        let row = (pos % (512*1792)) / 1792;
        let local = pos % 1792;
        let inner = (local/32)*64+part*32+local%32;
        let mut index = row*input.stride(0)+inner*input.stride(1);
        if comptime!(rhs) { index = inner*input.stride(0)+row*input.stride(1); }
        output[pos] = input[index];
    }
}

#[ruda(launch)]
fn reduce<F: Float>(partial: &Array<f32>, bias: &Tensor<F>, out: &mut Array<F>, #[define(F)] _dtype: StorageType) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() {
        let offset = (pos/512)*bias.stride(0)+(pos%512)*bias.stride(1);
        out[pos] = F::cast_from((partial[pos]+f32::cast_from(bias[offset]))+partial[512*512+pos]);
    }
}