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
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
impl OwnedQuantizedModelCachedSync {
/// Create a new thread-safe cached model wrapper
///
/// The scheduler is lazily initialized on first GPU operation.
/// The dequantized weight cache is lazily initialized via `warmup_gpu_cache()`.
/// PARITY-103: Also initializes CudaScheduler when CUDA feature is enabled.
#[must_use]
pub fn new(model: OwnedQuantizedModel) -> Self {
Self {
model,
scheduler: std::sync::Mutex::new(None),
#[cfg(feature = "cuda")]
cuda_scheduler: std::sync::Mutex::new(None),
dequant_cache: std::sync::RwLock::new(None),
}
}
/// Get reference to inner model
#[must_use]
pub fn model(&self) -> &OwnedQuantizedModel {
&self.model
}
/// Get or create the cached scheduler (thread-safe)
///
/// # Errors
/// Returns error if scheduler creation fails or lock is poisoned
fn get_scheduler(
&self,
) -> Result<std::sync::MutexGuard<'_, Option<crate::gpu::HybridScheduler>>> {
let mut scheduler_opt =
self.scheduler
.lock()
.map_err(|_| RealizarError::UnsupportedOperation {
operation: "scheduler_lock".to_string(),
reason: "Scheduler mutex poisoned".to_string(),
})?;
// Initialize if not already done
if scheduler_opt.is_none() {
use crate::gpu::HybridScheduler;
let new_scheduler = HybridScheduler::with_threshold(1000).map_err(|e| {
RealizarError::UnsupportedOperation {
operation: "HybridScheduler::with_threshold".to_string(),
reason: format!("GPU scheduler initialization failed: {e}"),
}
})?;
*scheduler_opt = Some(new_scheduler);
}
Ok(scheduler_opt)
}
/// PARITY-103: Get or create the cached CUDA scheduler (thread-safe)
///
/// Bypasses wgpu 256MB buffer limit by using cuBLAS directly.
/// Returns None if CUDA is not available.
///
/// # Errors
/// Returns error if lock is poisoned
#[cfg(feature = "cuda")]
fn get_cuda_scheduler(
&self,
) -> Result<std::sync::MutexGuard<'_, Option<crate::gpu::CudaScheduler>>> {
use crate::gpu::CudaScheduler;
let mut scheduler_opt =
self.cuda_scheduler
.lock()
.map_err(|_| RealizarError::UnsupportedOperation {
operation: "cuda_scheduler_lock".to_string(),
reason: "CUDA scheduler mutex poisoned".to_string(),
})?;
// Initialize if not already done
if scheduler_opt.is_none() {
match CudaScheduler::new() {
Ok(new_scheduler) => {
eprintln!("PARITY-103: CudaScheduler initialized successfully");
*scheduler_opt = Some(new_scheduler);
},
Err(e) => {
// CUDA not available, leave as None (will fallback to wgpu)
eprintln!("PARITY-103: CudaScheduler::new() failed: {:?}", e);
},
}
}
Ok(scheduler_opt)
}
/// PARITY-103: Batch matmul preferring CUDA over wgpu (thread-safe)
///
/// Tries CudaScheduler first (no buffer limits), falls back to HybridScheduler (wgpu).
/// This bypasses the wgpu 256MB buffer limit that was blocking GPU batch inference.
#[cfg(feature = "cuda")]
fn batch_matmul_gpu_prefer_cuda(
&self,
input: &[f32],
weight_f32: &[f32],
batch_size: usize,
in_dim: usize,
out_dim: usize,
) -> Result<Vec<f32>> {
// Validate input
if input.len() != batch_size * in_dim {
return Err(RealizarError::InvalidShape {
reason: format!(
"Input size {} doesn't match batch_size={} * in_dim={}",
input.len(),
batch_size,
in_dim
),
});
}
// Try CUDA first (no buffer size limits)
if let Ok(mut cuda_guard) = self.get_cuda_scheduler() {
if let Some(ref mut cuda_sched) = *cuda_guard {
return cuda_sched
.matmul(input, weight_f32, batch_size, in_dim, out_dim)
.map_err(|e| RealizarError::UnsupportedOperation {
operation: "batch_matmul_gpu_prefer_cuda".to_string(),
reason: format!("CUDA matmul failed: {e}"),
});
}
}
// Fallback to wgpu (may hit 256MB limit for large batches)
let mut scheduler_guard = self.get_scheduler()?;
if let Some(ref mut scheduler) = *scheduler_guard {
return scheduler
.matmul(input, weight_f32, batch_size, in_dim, out_dim)
.map_err(|e| RealizarError::UnsupportedOperation {
operation: "batch_matmul_gpu_prefer_cuda".to_string(),
reason: format!("GPU matmul failed: {e}"),
});
}
Err(RealizarError::UnsupportedOperation {
operation: "batch_matmul_gpu_prefer_cuda".to_string(),
reason: "No GPU scheduler available".to_string(),
})
}
/// PARITY-103: Batch matmul preferring CUDA (non-CUDA fallback)
#[cfg(not(feature = "cuda"))]
fn batch_matmul_gpu_prefer_cuda(
&self,
input: &[f32],
weight_f32: &[f32],
batch_size: usize,
in_dim: usize,
out_dim: usize,
) -> Result<Vec<f32>> {
// Validate input
if input.len() != batch_size * in_dim {
return Err(RealizarError::InvalidShape {
reason: format!(
"Input size {} doesn't match batch_size={} * in_dim={}",
input.len(),
batch_size,
in_dim
),
});
}
let mut scheduler_guard = self.get_scheduler()?;
if let Some(ref mut scheduler) = *scheduler_guard {
return scheduler
.matmul(input, weight_f32, batch_size, in_dim, out_dim)
.map_err(|e| RealizarError::UnsupportedOperation {
operation: "batch_matmul_gpu_prefer_cuda".to_string(),
reason: format!("GPU matmul failed: {e}"),
});
}
Err(RealizarError::UnsupportedOperation {
operation: "batch_matmul_gpu_prefer_cuda".to_string(),
reason: "No GPU scheduler available".to_string(),
})
}
/// Generate tokens with KV cache using thread-safe cached scheduler
///
/// Delegates to the inner model's `generate_with_cache` method.
/// The scheduler caching benefits GPU batch operations; single-token
/// generation uses CPU path with KV cache for O(n) scaling.
///
/// # Arguments
/// * `prompt` - Input token IDs
/// * `config` - Generation configuration
///
/// # Returns
/// Generated token sequence including prompt
///
/// # Errors
/// Returns error if generation fails
pub fn generate_with_cache(
&self,
prompt: &[u32],
config: &QuantizedGenerateConfig,
) -> Result<Vec<u32>> {
// Delegate to inner model - CPU path with KV cache is already efficient
self.model.generate_with_cache(prompt, config)
}
/// Generate tokens with adaptive CPU/GPU attention (IMP-126)
///
/// This variant of `generate_with_cache` uses adaptive CPU/GPU dispatch
/// based on cache length and records dispatch decisions to metrics.
///
/// # Arguments
/// * `prompt` - Initial token IDs
/// * `config` - Generation configuration
/// * `metrics` - Dispatch metrics tracker for CPU/GPU decision recording
///
/// # Returns
/// Generated token sequence including prompt
///
/// # Errors
/// Returns error if generation fails
#[cfg(feature = "gpu")]
pub fn generate_with_cache_adaptive(
&self,
prompt: &[u32],
config: &QuantizedGenerateConfig,
metrics: &std::sync::Arc<DispatchMetrics>,
) -> Result<Vec<u32>> {
// Delegate to inner model's adaptive generation
self.model
.generate_with_cache_adaptive(prompt, config, metrics)
}
/// Forward pass with cached scheduler (thread-safe)
///
/// Uses the cached HybridScheduler for GPU operations.
///
/// # Errors
/// Returns error if GPU operations fail
#[allow(clippy::let_underscore_untyped)] // Placeholder for future use
pub fn forward_batch_gpu_cached(&self, token_ids: &[u32]) -> Result<Vec<f32>> {
let batch_size = token_ids.len();
let vocab_size = self.model.config.vocab_size;
// Get cached scheduler (for future GPU operations)
let mut scheduler_guard = self.get_scheduler()?;
let _ = scheduler_guard
.as_mut()
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "forward_batch_gpu_cached".to_string(),
reason: "Scheduler not initialized".to_string(),
})?;
// 1. Token embedding lookup
let hidden = self.model.embed(token_ids);
// 2. Process through layers
for layer in &self.model.layers {
// Simplified single-layer forward - reuse inner model logic
// For full implementation, would need to port the complete forward pass
let _ = layer;
}
// 3. Output normalization and LM head
// For now, return placeholder - full implementation requires porting forward logic
let output = vec![0.0f32; batch_size * vocab_size];
let _ = hidden;
Ok(output)
}
/// Adaptive fused attention for production serving (IMP-121)
///
/// Thread-safe wrapper that automatically selects CPU or GPU based on
/// sequence length. Uses the cached scheduler for efficient GPU operations.
///
/// # Arguments
/// * `q` - Query tensor [seq_len, head_dim]
/// * `k` - Key tensor [seq_len, head_dim]
/// * `v` - Value tensor [seq_len, head_dim]
/// * `seq_len` - Sequence length
/// * `head_dim` - Head dimension
/// * `scale` - Attention scale factor
///
/// # Returns
/// Output tensor [seq_len, head_dim]
pub fn adaptive_fused_attention(
&self,
q: &[f32],
k: &[f32],
v: &[f32],
seq_len: usize,
head_dim: usize,
scale: f32,
) -> Result<Vec<f32>> {
// Threshold for GPU dispatch (from IMP-119 analysis)
const GPU_SEQ_LEN_THRESHOLD: usize = 64;
if seq_len >= GPU_SEQ_LEN_THRESHOLD {
// Long sequence: Use GPU path
self.gpu_fused_causal_attention(q, k, v, seq_len, head_dim, scale)
} else {
// Short sequence: Use CPU path
self.cpu_fused_causal_attention(q, k, v, seq_len, head_dim, scale)
}
}
/// CPU fused causal attention (thread-safe wrapper)
fn cpu_fused_causal_attention(
&self,
q: &[f32],
k: &[f32],
v: &[f32],
seq_len: usize,
head_dim: usize,
scale: f32,
) -> Result<Vec<f32>> {
// Use tiled implementation from inner model
self.model
.tiled_causal_attention(q, k, v, seq_len, head_dim, scale, 4)
}
/// GPU fused causal attention (thread-safe)
fn gpu_fused_causal_attention(
&self,
q: &[f32],
k: &[f32],
v: &[f32],
seq_len: usize,
head_dim: usize,
scale: f32,
) -> Result<Vec<f32>> {
let mut scheduler_guard =
self.scheduler
.lock()
.map_err(|_| RealizarError::UnsupportedOperation {
operation: "gpu_fused_causal_attention".to_string(),
reason: "Failed to acquire scheduler lock".to_string(),
})?;
// Initialize scheduler if needed
if scheduler_guard.is_none() {
use crate::gpu::HybridScheduler;
let new_scheduler = HybridScheduler::with_threshold(1000).map_err(|e| {
RealizarError::UnsupportedOperation {
operation: "HybridScheduler::with_threshold".to_string(),
reason: format!("GPU scheduler initialization failed: {e}"),
}
})?;
*scheduler_guard = Some(new_scheduler);
}
let scheduler =
scheduler_guard
.as_mut()
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "gpu_fused_causal_attention".to_string(),
reason: "Scheduler not initialized".to_string(),
})?;
// Transpose K for matmul
let mut k_transposed = vec![0.0f32; head_dim * seq_len];
for pos in 0..seq_len {
for d in 0..head_dim {
k_transposed[d * seq_len + pos] = k[pos * head_dim + d];
}
}
// GPU Q @ K^T
let scores = scheduler
.matmul(q, &k_transposed, seq_len, head_dim, seq_len)
.map_err(|e| RealizarError::UnsupportedOperation {
operation: "gpu_fused Q@K^T".to_string(),
reason: format!("GPU matmul failed: {}", e),
})?;
// CPU causal softmax
let mut weights = vec![0.0f32; seq_len * seq_len];
for i in 0..seq_len {
let mut max_val = f32::NEG_INFINITY;
for j in 0..=i {
let score = scores[i * seq_len + j] * scale;
if score > max_val {
max_val = score;
}
}
let mut sum = 0.0f32;
for j in 0..=i {
let score = scores[i * seq_len + j] * scale;
weights[i * seq_len + j] = (score - max_val).exp();
sum += weights[i * seq_len + j];
}
if sum > 0.0 {
for j in 0..=i {
weights[i * seq_len + j] /= sum;
}
}
}
// GPU weights @ V
scheduler
.matmul(&weights, v, seq_len, seq_len, head_dim)
.map_err(|e| RealizarError::UnsupportedOperation {
operation: "gpu_fused weights@V".to_string(),
reason: format!("GPU matmul failed: {}", e),
})
}
}