entrenar 0.7.9

Training & Optimization library with autograd, LoRA, quantization, and model merging
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

//! FP16 cuBLAS GEMM operations for training
//!
//! Contract: fp16-cublas-gemm-v1.yaml (PMAT-458)
//!
//! FP16 GEMM uses tensor cores on sm_89+ (83 TFLOPS vs 2 TFLOPS SIMD).
//! All matrices are FP16 (CUDA_R_16F) with FP32 accumulation (CUBLAS_COMPUTE_32F).
//! Expected: ~2x throughput vs fp32 on memory-BW-bound workloads.
//!
//! # Safety
//!
//! Backward GEMMs use `gemm_f16()` which internally uses CUBLAS_COMPUTE_32F
//! (FP32 accumulation). This is safe for transposed backward GEMMs — unlike
//! TF32 tensor cores which produce NaN at gradient magnitude ~1e5 (ALB-076).

#![allow(unsafe_code)]
#![allow(trivial_casts)]
#![allow(clippy::borrow_as_ptr)]
#![allow(clippy::ref_as_ptr)]

#[cfg(feature = "cuda")]
use trueno_gpu::driver::{CublasHandle, CudaStream, GemmOp, GpuBuffer};

use crate::autograd::cuda_tensor::{CudaTensorError, Result};

#[cfg(feature = "cuda")]
use super::cache::FORWARD_KERNEL_CACHE;

/// FP16 cuBLAS GEMM forward: C[M,N] = A[M,K] @ B[K,N] using tensor cores
///
/// Contract: fp16-cublas-gemm-v1.yaml C-FP16GEMM-001 (PMAT-458)
/// All matrices are FP16 (CUDA_R_16F). Accumulation in FP32 (CUBLAS_COMPUTE_32F).
/// Tensor cores activated via CUBLAS_GEMM_DEFAULT_TENSOR_OP.
/// Expected: ~2x throughput vs fp32 on memory-BW-bound workloads (RTX 4060L).
#[cfg(feature = "cuda")]
pub fn gemm_forward_f16(
    a: &GpuBuffer<u16>,
    b: &GpuBuffer<u16>,
    c: &mut GpuBuffer<u16>,
    m: u32,
    k: u32,
    n: u32,
    stream: &CudaStream,
) -> Result<()> {
    let cache = FORWARD_KERNEL_CACHE.get().ok_or(CudaTensorError::DeviceNotInitialized)?;
    let cache = cache.lock().map_err(|_err| {
        CudaTensorError::KernelError("Failed to acquire kernel cache lock".to_string())
    })?;
    let cublas = cache.cublas().ok_or_else(|| {
        CudaTensorError::KernelError("cuBLAS handle required for fp16 GEMM".to_string())
    })?;
    let _ = stream; // cuBLAS handle already bound to stream
    cublas
        .gemm_f16(
            GemmOp::NoTrans,
            GemmOp::NoTrans,
            n as i32,
            m as i32,
            k as i32,
            1.0,
            b.as_ptr(),
            n as i32,
            a.as_ptr(),
            k as i32,
            0.0,
            c.as_ptr(),
            n as i32,
        )
        .map_err(|e| {
            CudaTensorError::KernelError(format!("cuBLAS fp16 GEMM forward failed: {e:?}"))
        })
}

/// FP16 cuBLAS backward A: grad_A[M,K] = grad_C[M,N] @ B[K,N]^T (tensor cores)
///
/// Contract: fp16-cublas-gemm-v1.yaml C-FP16GEMM-002 (PMAT-458)
/// Gradient GEMM uses fp16 for memory bandwidth savings. Gradient accumulation
/// should be promoted to fp32 in the caller to prevent underflow.
/// Note: trueno gemm_f16 uses CUBLAS_COMPUTE_32F (fp32 accumulation), which
/// is safe for transposed backward GEMMs (unlike TF32 per ALB-076).
#[cfg(feature = "cuda")]
pub(crate) fn cublas_gemm_backward_a_f16(
    cublas: &CublasHandle,
    grad_output: &GpuBuffer<u16>,
    b: &GpuBuffer<u16>,
    grad_a: &mut GpuBuffer<u16>,
    m: u32,
    k: u32,
    n: u32,
) -> Result<()> {
    cublas
        .gemm_f16(
            GemmOp::Trans,
            GemmOp::NoTrans,
            k as i32,
            m as i32,
            n as i32,
            1.0,
            b.as_ptr(),
            n as i32,
            grad_output.as_ptr(),
            n as i32,
            0.0,
            grad_a.as_ptr(),
            k as i32,
        )
        .map_err(|e| CudaTensorError::KernelError(format!("cuBLAS fp16 backward_a failed: {e:?}")))
}

/// FP16 cuBLAS backward B: grad_B[K,N] = A[M,K]^T @ grad_C[M,N] (tensor cores)
///
/// Contract: fp16-cublas-gemm-v1.yaml C-FP16GEMM-002 (PMAT-458)
#[cfg(feature = "cuda")]
pub(crate) fn cublas_gemm_backward_b_f16(
    cublas: &CublasHandle,
    a: &GpuBuffer<u16>,
    grad_output: &GpuBuffer<u16>,
    grad_b: &mut GpuBuffer<u16>,
    m: u32,
    k: u32,
    n: u32,
) -> Result<()> {
    cublas
        .gemm_f16(
            GemmOp::NoTrans,
            GemmOp::Trans,
            n as i32,
            k as i32,
            m as i32,
            1.0,
            grad_output.as_ptr(),
            n as i32,
            a.as_ptr(),
            k as i32,
            0.0,
            grad_b.as_ptr(),
            n as i32,
        )
        .map_err(|e| CudaTensorError::KernelError(format!("cuBLAS fp16 backward_b failed: {e:?}")))
}

/// Mixed-precision backward_a: grad_A(fp32) = grad_C(fp16) @ B(fp16)^T (tensor cores)
///
/// Contract: C-FP16GEMM-002 (PMAT-472)
/// Enables dropping fp32 weights: backward uses fp16 weights with fp32 accumulation.
/// Cast grad_output fp32→fp16 at call site, pass fp16 weights, get fp32 grad_input.
#[cfg(feature = "cuda")]
pub fn gemm_f16_to_f32_backward_a(
    grad_output: &GpuBuffer<u16>,
    b: &GpuBuffer<u16>,
    grad_a: &mut GpuBuffer<f32>,
    m: u32,
    k: u32,
    n: u32,
    stream: &CudaStream,
) -> Result<()> {
    let cache = FORWARD_KERNEL_CACHE.get().ok_or(CudaTensorError::DeviceNotInitialized)?;
    let cache = cache.lock().map_err(|_err| {
        CudaTensorError::KernelError("Failed to acquire kernel cache lock".to_string())
    })?;
    let cublas = cache.cublas().ok_or_else(|| {
        CudaTensorError::KernelError("cuBLAS handle required for fp16→fp32 backward".to_string())
    })?;
    let _ = stream;
    cublas
        .gemm_f16_to_f32(
            GemmOp::Trans,
            GemmOp::NoTrans,
            k as i32,
            m as i32,
            n as i32,
            1.0,
            b.as_ptr(),
            n as i32,
            grad_output.as_ptr(),
            n as i32,
            0.0,
            grad_a.as_ptr(),
            k as i32,
        )
        .map_err(|e| {
            CudaTensorError::KernelError(format!("cuBLAS fp16→fp32 backward_a failed: {e:?}"))
        })
}

/// Mixed-precision GEMM: C(fp32) = A(fp16) @ B(fp16) using tensor cores
///
/// Contract: fp16-cublas-gemm-v1.yaml C-FP16GEMM-001 (PMAT-470)
/// A and B are FP16 (weights and activations cast to fp16). C is FP32.
/// Uses CUBLAS_COMPUTE_32F with CUBLAS_GEMM_DEFAULT_TENSOR_OP.
/// This is the "practical FP16 path": cast fp32 activations to fp16,
/// multiply by fp16 weights, produce fp32 output for the rest of the pipeline.
#[cfg(feature = "cuda")]
pub fn gemm_f16_to_f32_forward(
    a: &GpuBuffer<u16>,
    b: &GpuBuffer<u16>,
    c: &mut GpuBuffer<f32>,
    m: u32,
    k: u32,
    n: u32,
    stream: &CudaStream,
) -> Result<()> {
    let cache = FORWARD_KERNEL_CACHE.get().ok_or(CudaTensorError::DeviceNotInitialized)?;
    let cache = cache.lock().map_err(|_err| {
        CudaTensorError::KernelError("Failed to acquire kernel cache lock".to_string())
    })?;
    let cublas = cache.cublas().ok_or_else(|| {
        CudaTensorError::KernelError("cuBLAS handle required for fp16→fp32 GEMM".to_string())
    })?;
    let _ = stream;
    cublas
        .gemm_f16_to_f32(
            GemmOp::NoTrans,
            GemmOp::NoTrans,
            n as i32,
            m as i32,
            k as i32,
            1.0,
            b.as_ptr(),
            n as i32,
            a.as_ptr(),
            k as i32,
            0.0,
            c.as_ptr(),
            n as i32,
        )
        .map_err(|e| {
            CudaTensorError::KernelError(format!("cuBLAS fp16→fp32 GEMM forward failed: {e:?}"))
        })
}