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
use crate::ModelState;
use crate::modules::rope::RotaryEmbedding;
use candle_core::{DType, Result, Tensor};
use candle_nn::{Linear, Module, VarBuilder};
use std::collections::HashMap;
pub struct StreamingMultiheadAttention {
embed_dim: usize,
num_heads: usize,
rope: RotaryEmbedding,
in_proj: Linear,
out_proj: Linear,
context: Option<usize>,
name: String,
}
impl StreamingMultiheadAttention {
pub fn new(
embed_dim: usize,
num_heads: usize,
rope: RotaryEmbedding,
context: Option<usize>,
name: &str,
vb: VarBuilder,
) -> Result<Self> {
// out_dim = embed_dim + 2 * kv_dim (GQA/MHA logic in original)
// Original code:
// out_dim = embed_dim
// num_kv = num_heads
// kv_dim = (embed_dim // num_heads) * num_kv -> so embed_dim
// out_dim += 2 * kv_dim -> so 3 * embed_dim
let in_proj = candle_nn::linear_no_bias(embed_dim, 3 * embed_dim, vb.pp("in_proj"))?;
let out_proj = candle_nn::linear_no_bias(embed_dim, embed_dim, vb.pp("out_proj"))?;
Ok(Self {
embed_dim,
num_heads,
rope,
in_proj,
out_proj,
context,
name: name.to_string(),
})
}
pub fn init_state(
&self,
batch_size: usize,
sequence_length: usize,
device: &candle_core::Device,
) -> Result<HashMap<String, Tensor>> {
let dim_per_head = self.embed_dim / self.num_heads;
let mut state = HashMap::new();
state.insert(
"current_end_len".to_string(),
Tensor::zeros((1,), DType::U32, device)?,
);
state.insert(
"cache".to_string(),
Tensor::full(
f32::NAN,
(2, batch_size, sequence_length, self.num_heads, dim_per_head),
device,
)?,
);
Ok(state)
}
pub fn forward(&self, query: &Tensor, model_state: &mut ModelState) -> Result<Tensor> {
let projected = self.in_proj.forward(query)?;
let (b, t, _) = projected.dims3()?;
let d = self.embed_dim / self.num_heads;
// Auto-initialize state if missing
if !model_state.contains_key(&self.name) {
// Heuristic for KV cache size:
// If t is small (generation/streaming), reserve space for future tokens (e.g. 100x).
// If t is large (prompt processing), reserve just enough plus a small buffer.
// This prevents allocating 100x memory for long audio prompts (e.g., 100 * 10MB = 1GB).
let capacity = if t > 100 {
t + 2048 // Prompt + reasonable generation buffer
} else {
t * 100 // Short start, expect generation
};
let init = self.init_state(b, capacity, query.device())?;
model_state.insert(self.name.clone(), init);
}
let module_state = model_state.get_mut(&self.name).unwrap();
// Reshape to (b, t, 3, h, d)
let packed = projected.reshape((b, t, 3, self.num_heads, d))?;
let q = packed.narrow(2, 0, 1)?.squeeze(2)?;
let k = packed.narrow(2, 1, 1)?.squeeze(2)?;
let v = packed.narrow(2, 2, 1)?.squeeze(2)?;
let current_end = module_state
.get("current_end_len")
.ok_or_else(|| candle_core::Error::Msg("current_end_len not found".to_string()))?
.to_vec1::<u32>()?[0] as usize;
let (q, k) = self.rope.forward(&q, &k, current_end)?;
// Update KV cache
let _cache = module_state.get_mut("cache").unwrap();
// cache is (2, B, S, H, D)
// k, v are (B, T, H, D)
// We need to copy k to cache[0, :, current_end:current_end+T, :, :]
// and v to cache[1, :, current_end:current_end+T, :, :]
// This is tricky in Candle without in-place mutation of a tensor that is shared.
// For now, we'll use a simplified implementation where we slice and concat or narrow.
// However, if we want actual performance and correctness, we need to manage this cache.
// Let's implement a simplified KV update for now (concatenation) and optimize later.
// But wait, the original code uses a pre-allocated cache.
// To update a slice in Candle:
// We can't do it easily on a Tensor. We should probably store the cache as a list of chunks or
// use a single tensor and `index_copy` if available, or recreate it.
// Let's use `slice_assign` logic if it exists, or just concat for the prototype.
// Actually, Candle doesn't have slice_assign.
// Let's use the `current_end` to narrow the cache and then update it.
// But we want to avoid re-allocating the whole cache every step.
// For Phase 2, I'll use a growing KV cache (simple concat) to get the logic right.
let k_state = if current_end == 0 {
k.clone()
} else {
let k_prev = module_state
.get("k_cache")
.ok_or_else(|| candle_core::Error::Msg("k_cache not found".to_string()))?
.clone();
Tensor::cat(&[k_prev, k.clone()], 1)?
};
let v_state = if current_end == 0 {
v.clone()
} else {
let v_prev = module_state
.get("v_cache")
.ok_or_else(|| candle_core::Error::Msg("v_cache not found".to_string()))?
.clone();
Tensor::cat(&[v_prev, v.clone()], 1)?
};
// Compute attention using memory-efficient tiled implementation
let q_t = q.transpose(1, 2)?;
let k_t = k_state.transpose(1, 2)?;
let v_t = v_state.transpose(1, 2)?;
let scale = 1.0 / (d as f64).sqrt();
// Output: [B, H, T, D]
// We pass is_causal=true (since it's a streaming/causal model) and the context window.
// The sdpa function handles on-the-fly mask generation per tile.
let x = crate::modules::sdpa::sdpa(
&q_t,
&k_t,
&v_t,
scale,
true, // is_causal
self.context, // context_window
)?;
// Transpose back to [B, T, H, D] for output projection
// let x = x.transpose(1, 2)?.reshape((b, t, self.embed_dim))?; -- this is done in next lines usually
let x = x.transpose(1, 2)?.reshape((b, t, self.embed_dim))?;
let x = self.out_proj.forward(&x)?;
// Update state
module_state.insert("k_cache".to_string(), k_state);
module_state.insert("v_cache".to_string(), v_state);
module_state.insert(
"current_end_len".to_string(),
Tensor::from_vec(vec![(current_end + t) as u32], (1,), q.device())?,
);
Ok(x)
}
}