realizar 0.8.5

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370

impl CudaExecutor {
    // =========================================================================
    // PAR-023: Activation and Element-wise GPU Operations
    // =========================================================================

    /// PAR-023: SiLU activation on GPU buffer
    ///
    /// Computes: output[i] = input[i] * sigmoid(input[i])
    ///
    /// # Returns
    ///
    /// GPU buffer with activated result (no sync - async)
    pub fn silu_gpu(&mut self, input: &GpuBuffer<f32>, n: u32) -> Result<GpuBuffer<f32>, GpuError> {
        let kernel_type = KernelType::Silu { n };
        let kernel_name = self.kernels.kernel_name(&kernel_type);
        let cache_key = format!("silu_{}", n);

        if !self.modules.contains_key(&cache_key) {
            let ptx = self.kernels.generate_ptx(&kernel_type);
            let module = self.compile_ptx(&ptx)?;
            self.modules.insert(cache_key.clone(), module);
        }

        let module = self
            .modules
            .get_mut(&cache_key)
            .expect("module just inserted");

        let output = GpuBuffer::<f32>::new(&self.context, n as usize)?;

        // 256 threads per block for element-wise ops
        let threads = 256;
        let blocks = (n + threads - 1) / threads;
        let config = LaunchConfig::grid_2d(blocks, 1, threads, 1);

        let mut ptr_input = input.as_ptr();
        let mut ptr_output = output.as_ptr();
        let mut n_val = n;

        // SAFETY: Memory safety ensured by bounds checking and alignment
        unsafe {
            self.stream.launch_kernel(
                module,
                kernel_name,
                &config,
                &mut [
                    std::ptr::from_mut(&mut ptr_input) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
                ],
            )?;
        }

        Ok(output)
    }

    /// PAR-023: GELU activation on GPU buffer (async, returns new buffer)
    ///
    /// Computes approximate GELU: 0.5 * x * (1 + tanh(sqrt(2/π) * (x + 0.044715 * x³)))
    ///
    /// Unlike `gelu_gpu`, this returns a new buffer for async pipeline use.
    ///
    /// # Returns
    ///
    /// GPU buffer with activated result (no sync - async)
    pub fn gelu_async(
        &mut self,
        input: &GpuBuffer<f32>,
        n: u32,
    ) -> Result<GpuBuffer<f32>, GpuError> {
        let kernel_type = KernelType::Gelu { n };
        let kernel_name = self.kernels.kernel_name(&kernel_type);
        let cache_key = format!("gelu_async_{}", n);

        if !self.modules.contains_key(&cache_key) {
            let ptx = self.kernels.generate_ptx(&kernel_type);
            let module = self.compile_ptx(&ptx)?;
            self.modules.insert(cache_key.clone(), module);
        }

        let module = self
            .modules
            .get_mut(&cache_key)
            .expect("module just inserted");

        let output = GpuBuffer::<f32>::new(&self.context, n as usize)?;

        let threads = 256;
        let blocks = (n + threads - 1) / threads;
        let config = LaunchConfig::grid_2d(blocks, 1, threads, 1);

        let mut ptr_input = input.as_ptr();
        let mut ptr_output = output.as_ptr();
        let mut n_val = n;

        // SAFETY: Memory safety ensured by bounds checking and alignment
        unsafe {
            self.stream.launch_kernel(
                module,
                kernel_name,
                &config,
                &mut [
                    std::ptr::from_mut(&mut ptr_input) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
                ],
            )?;
        }

        Ok(output)
    }

    /// PAR-023: Element-wise multiply on GPU buffers
    ///
    /// Computes: output[i] = input1[i] * input2[i]
    /// Used for gated activations in SwiGLU.
    ///
    /// # Returns
    ///
    /// GPU buffer with product (no sync - async)
    pub fn elementwise_mul_gpu(
        &mut self,
        input1: &GpuBuffer<f32>,
        input2: &GpuBuffer<f32>,
        n: u32,
    ) -> Result<GpuBuffer<f32>, GpuError> {
        let kernel_type = KernelType::ElementwiseMul { n };
        let kernel_name = self.kernels.kernel_name(&kernel_type);
        let cache_key = format!("elementwise_mul_{}", n);

        if !self.modules.contains_key(&cache_key) {
            let ptx = self.kernels.generate_ptx(&kernel_type);
            let module = self.compile_ptx(&ptx)?;
            self.modules.insert(cache_key.clone(), module);
        }

        let module = self
            .modules
            .get_mut(&cache_key)
            .expect("module just inserted");

        let output = GpuBuffer::<f32>::new(&self.context, n as usize)?;

        let threads = 256;
        let blocks = (n + threads - 1) / threads;
        let config = LaunchConfig::grid_2d(blocks, 1, threads, 1);

        let mut ptr_input1 = input1.as_ptr();
        let mut ptr_input2 = input2.as_ptr();
        let mut ptr_output = output.as_ptr();
        let mut n_val = n;

        // SAFETY: Memory safety ensured by bounds checking and alignment
        unsafe {
            self.stream.launch_kernel(
                module,
                kernel_name,
                &config,
                &mut [
                    std::ptr::from_mut(&mut ptr_input1) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_input2) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
                ],
            )?;
        }

        Ok(output)
    }

    /// PAR-023: Fused SwiGLU activation on GPU buffers
    ///
    /// Computes: output[i] = silu(gate[i]) * up[i]
    /// Combines SiLU activation and multiply in one memory pass.
    ///
    /// # Returns
    ///
    /// GPU buffer with activated result (no sync - async)
    pub fn fused_swiglu_gpu(
        &mut self,
        gate: &GpuBuffer<f32>,
        up: &GpuBuffer<f32>,
        n: u32,
    ) -> Result<GpuBuffer<f32>, GpuError> {
        let kernel_type = KernelType::FusedSwiglu { n };
        let kernel_name = self.kernels.kernel_name(&kernel_type);
        let cache_key = format!("fused_swiglu_{}", n);

        if !self.modules.contains_key(&cache_key) {
            let ptx = self.kernels.generate_ptx(&kernel_type);
            let module = self.compile_ptx(&ptx)?;
            self.modules.insert(cache_key.clone(), module);
        }

        let module = self
            .modules
            .get_mut(&cache_key)
            .expect("module just inserted");

        let output = GpuBuffer::<f32>::new(&self.context, n as usize)?;

        let threads = 256;
        let blocks = (n + threads - 1) / threads;
        let config = LaunchConfig::grid_2d(blocks, 1, threads, 1);

        let mut ptr_gate = gate.as_ptr();
        let mut ptr_up = up.as_ptr();
        let mut ptr_output = output.as_ptr();
        let mut n_val = n;

        // SAFETY: Memory safety ensured by bounds checking and alignment
        unsafe {
            self.stream.launch_kernel(
                module,
                kernel_name,
                &config,
                &mut [
                    std::ptr::from_mut(&mut ptr_gate) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_up) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
                ],
            )?;
        }

        Ok(output)
    }

    /// PAR-044: Fused SwiGLU into existing buffer (zero-allocation, async)
    #[inline]
    pub fn fused_swiglu_into(
        &mut self,
        gate: &GpuBuffer<f32>,
        up: &GpuBuffer<f32>,
        output: &GpuBuffer<f32>,
        n: u32,
    ) -> Result<(), GpuError> {
        let kernel_type = KernelType::FusedSwiglu { n };
        let kernel_name = self.kernels.kernel_name(&kernel_type);
        let cache_key = format!("fused_swiglu_{}", n);

        if !self.modules.contains_key(&cache_key) {
            let ptx = self.kernels.generate_ptx(&kernel_type);
            let module = self.compile_ptx(&ptx)?;
            self.modules.insert(cache_key.clone(), module);
        }

        let module = self
            .modules
            .get_mut(&cache_key)
            .expect("module just inserted");

        let threads = 256;
        let blocks = (n + threads - 1) / threads;
        let config = LaunchConfig::grid_2d(blocks, 1, threads, 1);

        let mut ptr_gate = gate.as_ptr();
        let mut ptr_up = up.as_ptr();
        let mut ptr_output = output.as_ptr();
        let mut n_val = n;

        // SAFETY: Memory safety ensured by bounds checking and alignment
        unsafe {
            self.stream.launch_kernel(
                module,
                kernel_name,
                &config,
                &mut [
                    std::ptr::from_mut(&mut ptr_gate) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_up) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_output) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut n_val) as *mut std::ffi::c_void,
                ],
            )?;
        }

        // trueno#243: Record kernel for manual graph construction
        if self.graph_recording {
            let module = self.modules.get_mut(&cache_key).expect("module exists");
            let func = module.get_function(kernel_name)?;
            self.graph_recorded_kernels.push(RecordedKernel {
                func: SendCUfunction(func),
                config,
                arg_data: vec![ptr_gate, ptr_up, ptr_output, n_val as u64],
            });
        }

        Ok(())
    }

    /// PMAT-PERF-009: Fused Q/K/V projection on GPU
    ///
    /// Computes Q, K, V projections in a single kernel launch.
    /// Reduces kernel launch overhead from 3 launches to 1.
    ///
    /// # Arguments
    ///
    /// * `x` - Input hidden state [hidden_size]
    /// * `w_q` - Query weight matrix [hidden_size, hidden_size]
    /// * `w_k` - Key weight matrix [hidden_size, kv_dim]
    /// * `w_v` - Value weight matrix [hidden_size, kv_dim]
    /// * `out_q` - Output Q buffer [hidden_size]
    /// * `out_k` - Output K buffer [kv_dim]
    /// * `out_v` - Output V buffer [kv_dim]
    /// * `hidden_size` - Hidden dimension
    /// * `kv_dim` - KV dimension (for GQA, may differ from hidden_size)
    pub fn fused_qkv_into(
        &mut self,
        x: &GpuBuffer<f32>,
        w_q: &GpuBuffer<f32>,
        w_k: &GpuBuffer<f32>,
        w_v: &GpuBuffer<f32>,
        out_q: &GpuBuffer<f32>,
        out_k: &GpuBuffer<f32>,
        out_v: &GpuBuffer<f32>,
        hidden_size: u32,
        kv_dim: u32,
    ) -> Result<(), GpuError> {
        let kernel_type = KernelType::FusedQKV {
            hidden_size,
            kv_dim,
        };
        let kernel_name = self.kernels.kernel_name(&kernel_type);
        let cache_key = format!("fused_qkv_{}_{}", hidden_size, kv_dim);

        if !self.modules.contains_key(&cache_key) {
            let ptx = self.kernels.generate_ptx(&kernel_type);
            let module = self.compile_ptx(&ptx)?;
            self.modules.insert(cache_key.clone(), module);
        }

        let module = self
            .modules
            .get_mut(&cache_key)
            .expect("module just inserted");

        // Grid: one block per output row (max of hidden_size for Q, kv_dim for K/V)
        // Block: 32 threads (one warp) per row
        let rows = hidden_size.max(kv_dim);
        let config = LaunchConfig::grid_2d(rows, 1, 32, 1);

        let mut ptr_x = x.as_ptr();
        let mut ptr_wq = w_q.as_ptr();
        let mut ptr_wk = w_k.as_ptr();
        let mut ptr_wv = w_v.as_ptr();
        let mut ptr_out_q = out_q.as_ptr();
        let mut ptr_out_k = out_k.as_ptr();
        let mut ptr_out_v = out_v.as_ptr();

        // SAFETY: Memory safety ensured by bounds checking and alignment
        unsafe {
            self.stream.launch_kernel(
                module,
                kernel_name,
                &config,
                &mut [
                    std::ptr::from_mut(&mut ptr_x) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_wq) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_wk) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_wv) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_out_q) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_out_k) as *mut std::ffi::c_void,
                    std::ptr::from_mut(&mut ptr_out_v) as *mut std::ffi::c_void,
                ],
            )?;
        }

        Ok(())
    }
}