Module: SmolLM2KVCuda

Defined in:
lib/toy/llm/engine/llama_kv_engine_cuda.rb

Class Method Summary collapse

Class Method Details

.decode_step(kv_cache, token_id, pos) ⇒ Object

Decode one new token at position ‘pos`. Returns the (1, vocab) logits Mat for the new position. If `kv_cache.trace_on` is set the rebuild path inserts taps and we dump stats before reading logits.



1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
 
# File 'lib/toy/llm/engine/llama_kv_engine_cuda.rb', line 1516

def self.decode_step(kv_cache, token_id, pos)
  TinyNNCuda.tnn_reset_for_rebuild(kv_cache.sess)
  step = kv_cache.build_decode_step(pos)
  TinyNNCuda.tnn_realize(kv_cache.sess, step.kv_step_logits)
  TinyNNCuda.upload_int_array(kv_cache.sess, step.t_token_id, [token_id])
  TinyNNCuda.upload_int_array(kv_cache.sess, step.t_pos,      [pos])
  TinyNNCuda.tnn_compute(kv_cache.sess)
  kv_cache.dump_trace
  TinyNNCuda.download_row_major(kv_cache.sess, step.kv_step_logits, 1, kv_cache.vocab_size)
end

.upload_from(kv_cache, model) ⇒ Object

Upload all Toy::SmolLM2 weights into a realized cache (+ zero-init the K/V buffers).



1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
 
# File 'lib/toy/llm/engine/llama_kv_engine_cuda.rb', line 1449

def self.upload_from(kv_cache, model)
  sess     = kv_cache.sess
  n        = kv_cache.n_layers
  n_heads  = kv_cache.n_heads
  n_kv     = kv_cache.n_kv
  d_model  = kv_cache.d_model
  d_head   = kv_cache.d_head
  max_T    = kv_cache.max_T

  TinyNNCuda.upload_row_major(sess, kv_cache.t_token_embed, model.token_embed.weight)
  TinyNNCuda.tnn_upload_from_float_array(sess, kv_cache.t_final_norm_gamma,
                                      model.final_norm.gamma, d_model)
  if kv_cache.has_untied_output
    TinyNNCuda.upload_row_major(sess, kv_cache.t_output, model.output_proj)
  end

  # P5.2: K and V share the same layout ne=[d_head, max_T] now,
  # so they share the same zero-init Mat.
  kv_zero = Mat.new(max_T, d_head)

  li = 0
  while li < n
    blk_n = model.stack[li]
    blk_f = kv_cache.kv_blocks_ffi[li]

    TinyNNCuda.tnn_upload_from_float_array(sess, blk_f.t_rn1_gamma, blk_n.rn1.gamma, d_model)
    TinyNNCuda.tnn_upload_from_float_array(sess, blk_f.t_rn2_gamma, blk_n.rn2.gamma, d_model)

    hq = 0
    while hq < n_heads
      TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_q[hq], blk_n.attn.w_q[hq])
      if kv_cache.has_qkv_bias
        TinyNNCuda.tnn_upload_from_float_array(sess, blk_f.t_b_q[hq], blk_n.attn.b_q[hq], d_head)
      end
      hq = hq + 1
    end

    hkv = 0
    while hkv < n_kv
      TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_k[hkv], blk_n.attn.w_k[hkv])
      TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_v[hkv], blk_n.attn.w_v[hkv])
      if kv_cache.has_qkv_bias
        TinyNNCuda.tnn_upload_from_float_array(sess, blk_f.t_b_k[hkv], blk_n.attn.b_k[hkv], d_head)
        TinyNNCuda.tnn_upload_from_float_array(sess, blk_f.t_b_v[hkv], blk_n.attn.b_v[hkv], d_head)
      end
      # P5.1+P5.2: same Q8 skip rule as realize_for_mmap.
      if kv_cache.kv_type_k != 8
        TinyNNCuda.upload_row_major(sess, blk_f.t_K[hkv], kv_zero)
      end
      if kv_cache.kv_type_v != 8
        TinyNNCuda.upload_row_major(sess, blk_f.t_V[hkv], kv_zero)
      end
      hkv = hkv + 1
    end

    TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_o,    blk_n.attn.w_o)
    TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_gate, blk_n.ffn.w_gate)
    TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_up,   blk_n.ffn.w_up)
    TinyNNCuda.stage_transposed_and_upload(sess, blk_f.t_w_down, blk_n.ffn.w_down)

    li = li + 1
  end
end