Class: ToyLMMetal

Inherits:
Object
  • Object
show all
Defined in:
lib/toy/models/transformer_lm_metal.rb

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(arch) ⇒ ToyLMMetal

Returns a new instance of ToyLMMetal.



34
35
36
37
38
39
40
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 34

def initialize(arch)
  @arch    = arch
  @max_T   = 256
  @kv      = nil
  @gguf_handle = nil
  @loaded  = false
end

Instance Attribute Details

#arch ⇒ Object (readonly)

Returns the value of attribute arch.



32
33
34
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 32

def arch
  @arch
end

#max_T ⇒ Object

Returns the value of attribute max_T.



32
33
34
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 32

def max_T
  @max_T
end

#tokenizer ⇒ Object (readonly)

Returns the value of attribute tokenizer.



32
33
34
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 32

def tokenizer
  @tokenizer
end

Instance Method Details

#decode_step(token_id, pos) ⇒ Object 



91
92
93
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 91

def decode_step(token_id, pos)
  SmolLM2KVMetal.decode_step(@kv, token_id, pos)
end

#generate(prompt_ids, n_new, sampler_config = nil) ⇒ Object 



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
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 95

def generate(prompt_ids, n_new, sampler_config = nil)
  if !@loaded
    puts "ToyLMMetal.generate: model not loaded; call .load(path) first"
    return prompt_ids
  end
  ids = []
  j = 0
  while j < prompt_ids.length
    ids.push(prompt_ids[j])
    j = j + 1
  end

  i = 0
  while i < prompt_ids.length
    decode_step(prompt_ids[i], i)
    i = i + 1
  end

  ctx = nil
  if sampler_config != nil
    ctx = SamplerContext.new(ids, sampler_config.seed)
  end

  n = 0
  while n < n_new
    pos = ids.length
    last_id = ids[pos - 1]
    logits = decode_step(last_id, pos)
    pick = -1
    if sampler_config == nil
      pick = Sampler.argmax(logits)
    else
      logits = Sampler.repetition_penalty(logits, ctx, sampler_config.rep_penalty)
      logits = Sampler.temperature(logits, sampler_config.temperature)
      logits = Sampler.top_k(logits, sampler_config.top_k)
      logits = Sampler.top_p(logits, sampler_config.top_p)
      pick   = Sampler.pick(logits, sampler_config, ctx)
      ctx.generated_ids.push(pick)
    end
    ids.push(pick)
    if pick == @arch.eos_id
      break
    end
    n = n + 1
  end
  ids
end

#load(path) ⇒ Object 



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
 
# File 'lib/toy/models/transformer_lm_metal.rb', line 46

def load(path)
  flags = GGUFLoad.detect_smollm2_flags(path)
  cfg   = SmolLM2ConfigLoader.read(path)

  # Metal: skip the mmap fast path. tnn_session_attach_weight_mmap
  # falls through to ggml_backend_cpu_buffer_from_ptr on Metal (no
  # public Metal buffer_from_ptr API), and the ggml-metal scheduler
  # crashes during compute when fed CPU-resident weight tensors as
  # kernel inputs. The copy-load path (realize_for + load_kv_cache_auto)
  # is bytes-bounded by the weight tensor size; fine for the 135M
  # smoke and OK up to a few GB. A real Metal mmap path would need
  # ggml-metal to expose buffer-from-bytes-no-copy, tracked as a
  # follow-up on issue #2.

  kv = SmolLM2KVFFICacheMetal.new
  # P5.1: KV_Q8=1 opts into Q8_0 storage for the K cache. See the CPU
  # mirror in lib/transformer_lm.rb for the full rationale.
  if (ENV["KV_Q8"] || "") == "1"
    kv.enable_kv_q8!
  end
  # P4.1: FLASH_ATTN=1 → ggml_flash_attn_ext for attention.
  if (ENV["FLASH_ATTN"] || "") == "1"
    kv.enable_flash_attn!
  end

  if flags.qk_norm
    # #76, fail loud (never mask): Metal only has the copy-load path
    # (realize_for), which has no QK-norm support — the gamma tensors
    # are only allocated on the mmap path. Decoding a QK-norm model
    # (Qwen3 / OLMoE) here would be silently degenerate, exactly the
    # CUDA #76 failure mode. Abort until a Metal mmap/QK-norm path
    # exists.
    puts "ToyLMMetal.load: " + path + " needs QK-norm, which the " +
         "Metal copy-load path cannot apply (no mmap path on Metal " +
         "yet, issue #2). Aborting rather than decode degenerate."
    exit 1
  end
  kv.realize_for(@max_T, cfg.d_model, cfg.d_ff, cfg.n_heads, cfg.n_kv,
                 cfg.n_layers, cfg.vocab, cfg.rope_base, cfg.rms_eps,
                 flags.untied, flags.qkv_bias)
  GGUFLoad.load_kv_cache_auto(kv, path)
  @kv = kv
  @loaded = true
end