Module: Toy::LLM::Run::TrainHybrid

Defined in:
lib/toy/run/train_hybrid.rb

Constant Summary collapse

VOCAB = 
16
DM = 
8
H = 
2
S_V =

H*S_V == DM

4
T = 
4
STEPS = 
16
EPS = 
1.0e-5

Class Method Summary collapse

Class Method Details

.attention_layer(sess, t_x, rn, wq, wk, wv, wo, eps) ⇒ Object

Inline single-head causal self-attention (no RoPE/GQA — minimal, trainable). Weights arrive as explicit handles. Returns x + Wo·ctx.



67
68
69
70
71
72
73
74
75
76
77
78
79
80
 
# File 'lib/toy/run/train_hybrid.rb', line 67

def self.attention_layer(sess, t_x, rn, wq, wk, wv, wo, eps)
  h = Toy::LLM::Primitives::RMSNorm.build(sess, t_x, rn, eps)
  q = TinyNN.tnn_matmul(sess, wq, h)            # [DM, T]
  k = TinyNN.tnn_matmul(sess, wk, h)            # [DM, T]
  v = TinyNN.tnn_matmul(sess, wv, h)            # [DM, T]
  scores = TinyNN.tnn_matmul(sess, k, q)        # [T_k, T_q]
  scaled = TinyNN.tnn_scale(sess, scores, 1.0 / Math.sqrt(DM.to_f))
  masked = TinyNN.tnn_diag_mask_inf(sess, scaled, 0)
  attn   = TinyNN.tnn_softmax(sess, masked)     # [T_k, T_q]
  v_t    = TinyNN.tnn_transpose(sess, v)        # [T, DM]
  ctx    = TinyNN.tnn_matmul(sess, v_t, attn)   # [DM, T_q]
  out    = TinyNN.tnn_matmul(sess, wo, ctx)     # [DM, T]
  TinyNN.tnn_add(sess, t_x, out)
end

.fillv(n, seed) ⇒ Object 



45
46
47
48
49
50
51
52
53
 
# File 'lib/toy/run/train_hybrid.rb', line 45

def self.fillv(n, seed)
  a = [0.0]; a.pop
  i = 0
  while i < n
    a.push(((((i + seed) * 1103515245 + 12345) % 1000) - 500).to_f * 0.001)
    i = i + 1
  end
  a
end

.reg1(sess, pp, pm, pv, n) ⇒ Object 



214
215
216
217
218
219
 
# File 'lib/toy/run/train_hybrid.rb', line 214

def self.reg1(sess, pp, pm, pv, n)
  w = TinyNN.tnn_input_1d_f32_persistent(sess, n)
  pp.push(w); pm.push(TinyNN.tnn_input_1d_f32_persistent(sess, n))
  pv.push(TinyNN.tnn_input_1d_f32_persistent(sess, n))
  w
end

.reg2(sess, pp, pm, pv, rows, cols) ⇒ Object 



221
222
223
224
225
226
 
# File 'lib/toy/run/train_hybrid.rb', line 221

def self.reg2(sess, pp, pm, pv, rows, cols)
  w = TinyNN.tnn_input_2d_f32_persistent(sess, rows, cols)
  pp.push(w); pm.push(TinyNN.tnn_input_2d_f32_persistent(sess, rows, cols))
  pv.push(TinyNN.tnn_input_2d_f32_persistent(sess, rows, cols))
  w
end

.run ⇒ Object 



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
 
# File 'lib/toy/run/train_hybrid.rb', line 82

def self.run
  sess = TinyNN.tnn_session_new(0)
  TinyNN.tnn_session_set_graph_capacity(sess, 262144)

  # Flat param arrays (uniform ptr) so opt_step never sees two block types.
  pp = [TinyNN.tnn_null_ptr]; pp.pop
  pm = [TinyNN.tnn_null_ptr]; pm.pop
  pv = [TinyNN.tnn_null_ptr]; pv.pop

  # reg2/reg1: alloc a weight + matching m/v, register, return the weight.
  embed  = reg2(sess, pp, pm, pv, VOCAB, DM)   # ne0=DM, ne1=VOCAB
  fnorm  = reg1(sess, pp, pm, pv, DM)
  # Attention layer weights.
  a_rn   = reg1(sess, pp, pm, pv, DM)
  a_wq   = reg2(sess, pp, pm, pv, DM, DM)
  a_wk   = reg2(sess, pp, pm, pv, DM, DM)
  a_wv   = reg2(sess, pp, pm, pv, DM, DM)
  a_wo   = reg2(sess, pp, pm, pv, DM, DM)

  # GDN layer (its own weights live in ft_weights/ft_m/ft_v; flatten in).
  gblk = Toy::LLM::Blocks::GDNBlock.new
  gblk.alloc_trainable_f32_weights!(sess, DM, S_V, H)
  bi = 0
  while bi < gblk.ft_weights.length
    pp.push(gblk.ft_weights[bi]); pm.push(gblk.ft_m[bi]); pv.push(gblk.ft_v[bi])
    bi = bi + 1
  end

  # set_param BEFORE finalize (load-bearing order).
  gi = 0
  while gi < pp.length
    TinyNN.tnn_set_param(pp[gi])
    gi = gi + 1
  end
  TinyNN.tnn_finalize_weights(sess)
  gblk.zero_state!(sess)

  # Init weights + zero moments.
  gi = 0
  while gi < pp.length
    n = TinyNN.tnn_tensor_nelements(pp[gi])
    TinyNN.tnn_upload_from_float_array(sess, pp[gi], fillv(n, gi * 7 + 1), n)
    TinyNN.tnn_zero_tensor(sess, pm[gi])
    TinyNN.tnn_zero_tensor(sess, pv[gi])
    gi = gi + 1
  end

  # Forward — per-layer INT-kind dispatch (the seam pattern).
  t_tok = TinyNN.tnn_input_1d_i32(sess, T)
  x = TinyNN.tnn_get_rows(sess, embed, t_tok)
  kinds = [Toy::LLM::Archs::LayerSpec::KIND_ATTENTION,
           Toy::LLM::Archs::LayerSpec::KIND_GDN]
  li = 0
  while li < kinds.length
    if kinds[li] == Toy::LLM::Archs::LayerSpec::KIND_ATTENTION
      x = attention_layer(sess, x, a_rn, a_wq, a_wk, a_wv, a_wo, EPS)
    else
      x = gblk.build_forward(sess, x, T, EPS)
    end
    li = li + 1
  end
  xf  = Toy::LLM::Primitives::RMSNorm.build(sess, x, fnorm, EPS)
  lgt = TinyNN.tnn_matmul(sess, embed, xf)         # [VOCAB, T] tied

  t_labels = TinyNN.tnn_input_2d_f32(sess, T, VOCAB)
  t_hp     = TinyNN.tnn_input_1d_f32(sess, 7)
  t_loss   = TinyNN.tnn_cross_entropy_loss(sess, lgt, t_labels)
  TinyNN.tnn_set_output(t_loss)
  TinyNN.tnn_set_loss(t_loss)

  TinyNN.tnn_build_forward_only(sess, t_loss)
  TinyNN.tnn_build_backward(sess)
  gj = 0
  while gj < pp.length
    tg = TinyNN.tnn_tensor_grad(sess, pp[gj])
    to = TinyNN.tnn_opt_step_adamw(sess, pp[gj], tg, pm[gj], pv[gj], t_hp)
    TinyNN.tnn_extend_backward_graph(sess, to)
    gj = gj + 1
  end
  TinyNN.tnn_pin_all_graph_b_nodes(sess)
  TinyNN.tnn_realize_backward(sess)

  ids = [1, 2, 3, 4]
  labels = zeros(VOCAB * T)
  tt = 0
  while tt < T
    tgt = (ids[tt] + 1) % VOCAB
    labels[tgt + VOCAB * tt] = 1.0
    tt = tt + 1
  end
  hp = [0.02, 0.9, 0.95, 1.0e-8, 0.0, 0.9, 0.95]

  first_loss = 0.0
  last_loss  = 0.0
  s = 0
  while s < STEPS
    if s == 0
      TinyNN.tnn_graph_reset(sess)
    else
      TinyNN.tnn_graph_reset_grads_only(sess)
    end
    TinyNN.upload_int_array(sess, t_tok, ids)
    TinyNN.tnn_upload_from_float_array(sess, t_labels, labels, VOCAB * T)
    TinyNN.tnn_upload_from_float_array(sess, t_hp, hp, 7)
    TinyNN.tnn_compute_backward(sess)
    TinyNN.tnn_download(sess, t_loss)
    lv = TinyNN.tnn_scratch_get(sess, 0)
    if s == 0
      first_loss = lv
    end
    last_loss = lv
    puts "step " + s.to_s + ": loss=" + lv.to_s
    s = s + 1
  end

  ok = true
  if first_loss != first_loss || last_loss != last_loss
    puts "FAIL: loss is NaN"
    ok = false
  end
  if last_loss >= first_loss - 0.05
    puts "FAIL: loss did not decrease (first=" + first_loss.to_s + " last=" + last_loss.to_s + ")"
    ok = false
  end
  if ok
    puts "HYBRID train smoke PASS: attention+GDN from-scratch hybrid trains — CE loss " +
         first_loss.to_s + " -> " + last_loss.to_s + " over " + STEPS.to_s + " steps"
  else
    puts "HYBRID train smoke FAIL"
  end
end

.zeros(n) ⇒ Object 



55
56
57
58
59
60
61
62
63
 
# File 'lib/toy/run/train_hybrid.rb', line 55

def self.zeros(n)
  a = [0.0]; a.pop
  i = 0
  while i < n
    a.push(0.0)
    i = i + 1
  end
  a
end