tract-cuda 0.23.8

Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
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

use cudarc::driver::LaunchConfig;
use tract_core::internal::*;
use tract_gpu::tensor::DeviceTensor;

use crate::context::{TractCudaStream, cuda_context};
use crate::kernels::launch_args::TractLaunchArgs;
use crate::kernels::{LibraryName, get_cuda_view};

#[derive(Debug, Clone)]
pub struct CudaCausalConv1dUpdate;

impl CudaCausalConv1dUpdate {
    /// Returns (batch, channels, kernel_width, s_len). Layout: input/output
    /// [b, C, S] (S may be folded away, e.g. [b, C]), state [b, C, K],
    /// weight [C, K].
    pub fn validate(
        &self,
        input: &DeviceTensor,
        weight: &DeviceTensor,
        state: &DeviceTensor,
    ) -> TractResult<(usize, usize, usize, usize)> {
        ensure!(input.datum_type() == DatumType::F16);
        ensure!(weight.datum_type() == DatumType::F16 && state.datum_type() == DatumType::F16);
        let kernel_width = *weight.shape().last().context("conv weight must have a kernel axis")?;
        ensure!(kernel_width == 4, "Qwen3.5 requires a four-tap convolution");
        ensure!(state.rank() >= 2, "conv state must be [b, C, K]");
        let batch = if state.rank() == 3 { state.shape()[0] } else { 1 };
        ensure!(state.len().is_multiple_of(batch * kernel_width));
        let channels = state.len() / (batch * kernel_width);
        ensure!(weight.len() == channels * kernel_width);
        ensure!(input.len().is_multiple_of(batch * channels), "conv input not [b, C, S]");
        // Both supported layouts ([b, C, S] and the folded [b, C]) carry an
        // explicit batch axis at shape()[0]; it must agree with state's
        // batch -- otherwise input.len() being a multiple of batch * channels
        // can pass by coincidence with a mismatched batch folded into a
        // wrong s_len.
        if input.rank() >= 2 {
            ensure!(
                input.shape()[0] == batch,
                "conv input batch {} != state batch {}",
                input.shape()[0],
                batch
            );
        }
        let s_len = input.len() / (batch * channels);
        Ok((batch, channels, kernel_width, s_len))
    }

    pub fn dispatch_eval(
        &self,
        stream: &TractCudaStream,
        input: &DeviceTensor,
        weight: &DeviceTensor,
        state: &DeviceTensor,
        output: &DeviceTensor,
        final_state: &DeviceTensor,
    ) -> TractResult<()> {
        let (batch, channels, kernel_width, s_len) = self.validate(input, weight, state)?;
        ensure!(output.shape() == input.shape() && output.datum_type() == DatumType::F16);
        ensure!(final_state.shape() == state.shape() && final_state.datum_type() == DatumType::F16);
        let function = cuda_context().load_pipeline(
            LibraryName::GdnRecurrent,
            "tract_causal_conv1d_update_f16".to_string(),
        )?;
        let i = get_cuda_view(input);
        let w = get_cuda_view(weight);
        let s = get_cuda_view(state);
        let o = get_cuda_view(output);
        let ns = get_cuda_view(final_state);
        let mut args = TractLaunchArgs::new(stream, &function);
        args.push_view(&i);
        args.push_view(&w);
        args.push_view(&s);
        args.push_view(&o);
        args.push_view(&ns);
        args.push_i32(channels);
        args.push_i32(kernel_width);
        args.push_i32(s_len);
        args.push_i32(batch);
        args.launch(LaunchConfig::for_num_elems((batch * channels) as u32))?;
        Ok(())
    }

    pub fn eval(
        &self,
        stream: &TractCudaStream,
        input: &DeviceTensor,
        weight: &DeviceTensor,
        state: &DeviceTensor,
    ) -> TractResult<(DeviceTensor, DeviceTensor)> {
        let output = DeviceTensor::uninitialized_dt(DatumType::F16, input.shape())?;
        let final_state = DeviceTensor::uninitialized_dt(DatumType::F16, state.shape())?;
        self.dispatch_eval(stream, input, weight, state, &output, &final_state)?;
        stream.synchronize()?;
        Ok((output, final_state))
    }
}

pub fn cuda_causal_conv1d_update_launch(
    input: &DeviceTensor,
    weight: &DeviceTensor,
    state: &DeviceTensor,
    output: &DeviceTensor,
    final_state: &DeviceTensor,
) -> TractResult<()> {
    crate::with_cuda_stream(|stream| {
        CudaCausalConv1dUpdate.dispatch_eval(stream, input, weight, state, output, final_state)
    })
}

crate::register_cuda_op!(
    tract_transformers::ops::causal_conv1d_update::CausalConv1dUpdate,
    |source, node, _op| {
        // The cuda kernel is f16-only; the op's layout is [b, C, S] and the
        // kernel loops over S and batch (port of the Metal kernel).
        let facts = source.node_input_facts(node.id)?;
        let rank_ok = facts[0].rank() == 3;
        let dts_ok = facts.iter().all(|f| f.datum_type == DatumType::F16);
        // The kernel loops over S and batch internally, so symbolic or
        // concrete values of either are fine.
        if !rank_ok || !dts_ok {
            return Ok(None);
        }
        Ok(Some(Box::new(tract_gpu::ops::causal_conv1d_update::GpuCausalConv1dUpdate {
            backend_name: "Cuda",
            dispatch: cuda_causal_conv1d_update_launch,
        })))
    }
);

#[cfg(test)]
mod tests {
    use super::*;
    use tract_gpu::tensor::IntoDevice;

    #[test]
    fn qwen35_conv_update_matches_cpu() -> TractResult<()> {
        crate::with_cuda_stream(|stream| {
            let channels = 6144usize;
            let input_f = (0..channels).map(|i| ((i % 17) as f32 - 8.0) / 32.0).collect::<Vec<_>>();
            let weight_f =
                (0..channels * 4).map(|i| ((i % 13) as f32 - 6.0) / 64.0).collect::<Vec<_>>();
            let state_f =
                (0..channels * 4).map(|i| ((i % 11) as f32 - 5.0) / 32.0).collect::<Vec<_>>();
            let cvt = |v: &[f32]| v.iter().copied().map(f16::from_f32).collect::<Vec<_>>();
            let input = Tensor::from_shape(&[1, channels], &cvt(&input_f))?.into_device()?;
            let weight = Tensor::from_shape(&[channels, 4], &cvt(&weight_f))?.into_device()?;
            let state = Tensor::from_shape(&[1, channels, 4], &cvt(&state_f))?.into_device()?;
            let (output, next) = CudaCausalConv1dUpdate.eval(stream, &input, &weight, &state)?;
            let output = output.to_host()?.into_tensor();
            let next = next.to_host()?.into_tensor();
            let got = unsafe { output.as_slice_unchecked::<f16>() };
            let got_state = unsafe { next.as_slice_unchecked::<f16>() };
            for c in 0..channels {
                let base = c * 4;
                assert_eq!(&got_state[base..base + 3], &cvt(&state_f[base + 1..base + 4]));
                assert_eq!(got_state[base + 3], f16::from_f32(input_f[c]));
                let sum = (0..3).map(|t| state_f[base + t + 1] * weight_f[base + t]).sum::<f32>()
                    + input_f[c] * weight_f[base + 3];
                assert_eq!(got[c], f16::from_f32(sum / (1.0 + (-sum).exp())));
            }
            Ok(())
        })
    }
}