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

/// GH-560: wgpu (Vulkan) batch inference state.
///
/// Toyota Way: lazy mmap for embedding (zero-copy), no model.clone().
/// Dequantized weights uploaded to GPU once, reused across all prompts.
/// Embedding lookup uses mmap'd model file — zero heap allocation.
///
/// Five-whys root cause (OOM, GH-560):
///   1. Why OOM? 56 GB peak during init
///   2. Why 56 GB? dequant_model_weights() called TWICE (28 GB each)
///   3. Why twice? First for upload loop, second for lm_head extraction
///   4. Why separate? into_iter() consumed the Vec, couldn't find lm_head after
///   5. Why not extract inline? Oversight — fixed by extracting during iteration
///
/// Five-whys root cause (garbage output, GH-560):
///   1. Why garbage? Attention produces near-uniform scores → random tokens
///   2. Why uniform? KV cache has max_seq zero-vectors before actual K/V data
///   3. Why zeros? generate() zeroed cache contents but didn't clear length
///   4. Why wrong length? Pre-allocated vec![0.0; max_seq * kv_dim] starts at full length
///   5. Why full length? forward_layer uses extend_from_slice + len()/kv_dim for seq_len
///   - Fix: Vec::with_capacity (empty, pre-alloc) + clear() (reset length to 0)
///
// Included from batch.rs via include!() — shares the same module scope.
//
/// Holds wgpu forward pass + CPU LM head for batch inference.
#[cfg(feature = "gpu")]
struct WgpuBatchState {
    fwd: trueno::backends::gpu::WgslForwardPass,
    model_path: std::path::PathBuf,
    hidden_dim: usize,
    num_layers: usize,
    num_kv_heads: usize,
    head_dim: usize,
    vocab_size: usize,
    eps: f32,
    lm_head_f32: Vec<f32>,
    output_norm: Vec<f32>,
    kv_caches: Vec<(Vec<f32>, Vec<f32>)>,
}

#[cfg(feature = "gpu")]
impl WgpuBatchState {
    /// Generate tokens from input using wgpu forward pass.
    /// Contract: output must match CPU forward pass (cosine >= 0.98).
    #[provable_contracts_macros::contract("gpu-multi-backend-parity-v1", equation = "multi_backend_parity")]
    fn generate(
        &mut self,
        input_tokens: &[u32],
        config: &crate::gguf::QuantizedGenerateConfig,
    ) -> std::result::Result<Vec<u32>, RealizarError> {
        let hd = self.hidden_dim;
        let kv_dim = self.num_kv_heads * self.head_dim;

        // Lazy mmap for embedding lookup — zero-copy, no 7.5 GB clone
        let mapped = crate::apr::MappedAprModel::from_path(&self.model_path)?;
        let embed_model = crate::gguf::OwnedQuantizedModel::from_apr(&mapped)?;

        // Reset KV caches — clear() resets length to 0 (capacity retained).
        // forward_layer uses extend_from_slice + len()/kv_dim to derive seq_len,
        // so length MUST be 0 at start. Zeroing contents preserves wrong length.
        for (k, v) in &mut self.kv_caches {
            k.clear();
            v.clear();
        }

        let mut output_tokens = input_tokens.to_vec();

        // PREFILL: Process all input tokens EXCEPT the last to build KV cache.
        if input_tokens.len() > 1 {
            for (pos, &tok) in input_tokens[..input_tokens.len()-1].iter().enumerate() {
                let mut hidden = embed_model.embed(&[tok]);
                for layer_idx in 0..self.num_layers {
                    let prefix = format!("layer.{layer_idx}");
                    let (ref mut kv_k, ref mut kv_v) = self.kv_caches[layer_idx];
                    self.fwd.forward_layer(
                        &mut hidden, &prefix, pos, kv_k, kv_v,
                    ).map_err(|e| RealizarError::InferenceError(format!("wgpu prefill L{layer_idx}: {e}")))?;
                }
            }
        }

        // DECODE: Generate new tokens. First step processes the last input token.
        for _step in 0..config.max_tokens {
            let token_id = *output_tokens.last().unwrap();
            let position = output_tokens.len() - 1;

            let mut hidden = embed_model.embed(&[token_id]);
            for layer_idx in 0..self.num_layers {
                let prefix = format!("layer.{layer_idx}");
                let (ref mut kv_k, ref mut kv_v) = self.kv_caches[layer_idx];
                self.fwd.forward_layer(
                    &mut hidden, &prefix, position, kv_k, kv_v,
                ).map_err(|e| RealizarError::InferenceError(format!("wgpu L{layer_idx}: {e}")))?;
            }

            // Output norm + LM head argmax (CPU, 4x unrolled)
            let sq_sum: f32 = hidden.iter().map(|x| x * x).sum();
            let rms = (sq_sum / hd as f32 + self.eps).sqrt();
            let normed: Vec<f32> = hidden.iter().zip(self.output_norm.iter())
                .map(|(x, g)| (x / rms) * g).collect();

            let mut best_idx = 0u32;
            let mut best_val = f32::NEG_INFINITY;
            for i in 0..self.vocab_size {
                let row = &self.lm_head_f32[i * hd..(i + 1) * hd];
                let mut sum = 0.0f32;
                let chunks = hd / 4;
                for c in 0..chunks {
                    let j = c * 4;
                    sum += row[j]*normed[j] + row[j+1]*normed[j+1]
                         + row[j+2]*normed[j+2] + row[j+3]*normed[j+3];
                }
                for j in (chunks*4)..hd { sum += row[j] * normed[j]; }
                if sum > best_val { best_val = sum; best_idx = i as u32; }
            }

            output_tokens.push(best_idx);
            if config.stop_tokens.contains(&best_idx) { break; }
        }

        Ok(output_tokens)
    }
}

/// GH-560: Try initializing wgpu batch state.
///
/// Single-pass dequant: extracts lm_head inline during weight upload iteration.
/// Previous code called dequant_model_weights() twice (56 GB peak → 28 GB peak).
#[cfg(feature = "gpu")]
#[provable_contracts_macros::contract("gpu-weight-residency-v1", equation = "pcie_overhead")]
fn try_init_wgpu_batch(
    model: &crate::gguf::OwnedQuantizedModel,
    config: &BatchInferenceConfig,
) -> Option<WgpuBatchState> {
    use crate::gpu::adapters::wgpu_adapter;
    use trueno::backends::gpu::GpuDevice;

    if !GpuDevice::is_available() {
        return None;
    }

    let gpu = GpuDevice::new().ok()?;

    if config.verbose {
        eprintln!("[batch] Backend: wgpu (Vulkan)");
    }

    let cfg = model.config();
    let hidden_dim = cfg.hidden_dim;
    let num_heads = cfg.num_heads;
    let num_kv_heads = cfg.num_kv_heads;
    let head_dim = hidden_dim / num_heads;
    let intermediate_dim = cfg.intermediate_dim;
    let kv_dim = num_kv_heads * head_dim;
    let num_layers = cfg.num_layers;

    let mut fwd = trueno::backends::gpu::WgslForwardPass::new(
        gpu.device, gpu.queue,
        hidden_dim, num_heads, num_kv_heads, head_dim, intermediate_dim,
    );

    // C-WGPU-Q4K-001: Upload raw Q4K bytes for projection weights (7x less VRAM).
    let raw_q4k = wgpu_adapter::raw_q4k_weights(model);
    let q4k_names: std::collections::HashSet<String> =
        raw_q4k.iter().map(|(n, _, _, _)| n.clone()).collect();
    for (name, data, _rows, _cols) in &raw_q4k {
        fwd.upload_q4k_weight(name, data);
    }

    // GH-560 FIX: Dequantize non-Q4K weights in ONE pass. Extract lm_head during iteration.
    let weights = wgpu_adapter::dequant_model_weights(model).ok()?;
    let mut lm_head_f32 = Vec::new();
    for (name, data, _rows, _cols) in weights {
        if name == "lm_head" {
            lm_head_f32 = data;
            continue;
        }
        if !q4k_names.contains(&name) {
            fwd.upload_weight(&name, &data);
        }
    }

    // Upload biases (small, F32)
    for (i, layer) in model.layers().iter().enumerate() {
        let prefix = format!("layer.{i}");
        if let Some(ref bias) = layer.qkv_bias {
            let q_dim = num_heads * head_dim;
            let kv_d = num_kv_heads * head_dim;
            if bias.len() >= q_dim + 2 * kv_d {
                fwd.upload_weight(&format!("{prefix}.q_bias"), &bias[..q_dim]);
                fwd.upload_weight(&format!("{prefix}.k_bias"), &bias[q_dim..q_dim+kv_d]);
                fwd.upload_weight(&format!("{prefix}.v_bias"), &bias[q_dim+kv_d..q_dim+2*kv_d]);
            }
        }
    }

    // Output norm (small: hidden_dim floats)
    let output_norm = model.output_norm_weight().to_vec();

    // Initialize GPU KV caches
    fwd.init_kv_cache(num_layers);

    // KV caches: start empty (length 0), pre-allocate capacity to avoid reallocs.
    // forward_layer appends via extend_from_slice and derives seq_len from len().
    let max_seq = config.max_tokens + 128;
    let kv_caches: Vec<(Vec<f32>, Vec<f32>)> = (0..num_layers)
        .map(|_| (Vec::with_capacity(max_seq * kv_dim), Vec::with_capacity(max_seq * kv_dim)))
        .collect();

    Some(WgpuBatchState {
        fwd,
        model_path: config.model_path.clone(),
        hidden_dim, num_layers, num_kv_heads, head_dim,
        vocab_size: cfg.vocab_size, eps: cfg.eps,
        lm_head_f32, output_norm, kv_caches,
    })
}