Module: ToyDescribeFlow

Defined in:
lib/toy/dev/toy_describe_flow.rb

Overview

tao#kv-describe-flow — DAG introspector over a built compute graph.

Walks the session’s primary graph (‘graph_a` — populated by tnn_build_forward_only or tnn_realize), discovers leaves by scanning each node’s src tensors for non-node references (= weights, inputs, constants), classifies every tensor by role, and emits three forms:

ToyDescribeFlow.text(sess)    → terminal-readable summary
ToyDescribeFlow.mermaid(sess) → mermaid flowchart for md reports
ToyDescribeFlow.json(sess)    → structured DAG for Tao's HTML

Acceptance (Tao spec):

- round-trips through JSON.parse
- text labels weight tensors as `weight`, intermediates as `ephemeral`

Spinel notes: avoids String interpolation, Hash<Ptr,…> (not supported), and Math.exp / Array<Hash> patterns. Uses the seed-and-pop idiom for typed-pointer arrays (‘[TinyNN.tnn_null_ptr]; .pop` per docs/spinel-tech-debt-audit-2026-05-25.md landmine #1).

Class Method Summary collapse

Class Method Details

.build_index(sess) ⇒ Object

Build the full ptr→index lookup. Returns parallel arrays (ptrs, is_node) — leaves first (is_node=false), then compute nodes (is_node=true). Pure Spinel-friendly arithmetic; no Hash<Ptr, Int>.



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
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 90

def self.build_index(sess)
  nodes = [TinyNN.tnn_null_ptr]; nodes.pop
  leaves = [TinyNN.tnn_null_ptr]; leaves.pop

  n_nodes = TinyNN.tnn_graph_n_nodes(sess)
  i = 0
  while i < n_nodes
    nodes.push(TinyNN.tnn_graph_node(sess, i))
    i = i + 1
  end

  # For each node's src, check whether it's already known as a node
  # OR already added to leaves; if neither, it's a fresh leaf.
  i = 0
  while i < n_nodes
    node = nodes[i]
    si = 0
    while si < 10
      src = TinyNN.tnn_tensor_src(node, si)
      if src == nil || src == TinyNN.tnn_null_ptr
        si = 10
      else
        # Already a node? Linear scan over `nodes`.
        in_nodes = false
        k = 0
        while k < nodes.length
          if nodes[k] == src
            in_nodes = true
            k = nodes.length
          else
            k = k + 1
          end
        end
        if !in_nodes
          # Already in leaves?
          in_leaves = false
          k = 0
          while k < leaves.length
            if leaves[k] == src
              in_leaves = true
              k = leaves.length
            else
              k = k + 1
            end
          end
          if !in_leaves
            leaves.push(src)
          end
        end
        si = si + 1
      end
    end
    i = i + 1
  end

  # Return (ptrs, n_leafs) where ptrs = leaves + nodes; first n_leafs
  # entries are leaves, the rest are compute nodes.
  ptrs = [TinyNN.tnn_null_ptr]; ptrs.pop
  i = 0; while i < leaves.length; ptrs.push(leaves[i]); i = i + 1; end
  i = 0; while i < nodes.length;  ptrs.push(nodes[i]);  i = i + 1; end
  [ptrs, leaves.length]
end

.card(sess) ⇒ Object

P1 — derive a Toy::Card from a realized session by walking its compute graph. Structural-exhaustive view (every op, every shape, every tensor name from the graph). Complements the textbook ‘algorithm` methods on stdlib archs; doesn’t replace them.

Returns a Toy::Card with:

  • PARAM leaves → add_param (name, shape, “”)

  • INPUT leaves → add_input (name, shape, “”)

  • each compute node → step_bind (var, “OP(src1, src2, …)”, shape)

  • the last OUTPUT-flagged node → step_return

Anonymous tensors (empty names) are labeled “_t<index>” so the rendered pseudocode stays unambiguous.



332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 332

def self.card(sess)
  c = Toy::Card.new("derived", "structural")

  # Direct-index access to avoid the Array<Array<mixed>> destructure
  # landmine. Same pattern as the existing text/json/mermaid emitters.
  pair    = build_index(sess)
  ptrs    = pair[0]
  n_leafs = pair[1]

  # P1 limitation: building per-node Step records via
  # Card#step_bind with FFI :str args triggers a Spinel poly-cascade
  # landmine that broadens Step's field types and crashes static
  # init. For now, we record counts as hypers so callers get a
  # useful Card (param count, input count, node count) without
  # full per-step detail. Restoring the full step list is a
  # follow-up — either restructure Card to accept loosely-typed
  # builders, or move Step construction to C-side primitives
  # (`tnn_card_push_step`).
  n_param_leafs = 0
  n_input_leafs = 0
  j = 0
  while j < n_leafs
    lflags = TinyNN.tnn_tensor_flags(ptrs[j])
    if (lflags & 4) != 0; n_param_leafs = n_param_leafs + 1; end
    if (lflags & 1) != 0; n_input_leafs = n_input_leafs + 1; end
    j = j + 1
  end
  c.add_hyper("leafs.total",  n_leafs.to_s)
  c.add_hyper("leafs.params", n_param_leafs.to_s)
  c.add_hyper("leafs.inputs", n_input_leafs.to_s)
  c.add_hyper("nodes.compute", (ptrs.length - n_leafs).to_s)

  # Identify the OUTPUT — last OUTPUT-flagged node wins; fallback
  # to the final node.
  out_name = ""
  i = n_leafs
  while i < ptrs.length
    t     = ptrs[i]
    flags = TinyNN.tnn_tensor_flags(t)
    if (flags & 2) != 0
      name = TinyNN.tnn_tensor_name(t)
      if name == ""
        name = "_t" + i.to_s
      end
      out_name = name
    end
    i = i + 1
  end
  if out_name == "" && ptrs.length > n_leafs
    t = ptrs[ptrs.length - 1]
    name = TinyNN.tnn_tensor_name(t)
    if name == ""
      name = "_t" + (ptrs.length - 1).to_s
    end
    out_name = name
  end
  if out_name != ""
    c.step_return(out_name)
  end

  c
end

.dtype_name(t) ⇒ Object

ggml type enum (subset toy uses) → human name. Keep in sync with vendor/ggml/include/ggml.h GGML_TYPE_*.



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
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 24

def self.dtype_name(t)
  if t == 0
    "f32"
  elsif t == 1
    "f16"
  elsif t == 8
    "q8_0"
  elsif t == 9
    "q8_1"
  elsif t == 15
    "q8_k"
  elsif t == 24
    "i8"
  elsif t == 25
    "i16"
  elsif t == 26
    "i32"
  elsif t == 27
    "i64"
  elsif t == 30
    "bf16"
  else
    "type" + t.to_s
  end
end

.emit_flow_json(run_dir, sess) ⇒ Object

tao#flow-json-emit (issue #25): write the realized graph’s flow.json into ‘run_dir` (a no-op when run_dir is “”). Called by every training runner right after realize! so the run bundle is self-describing — parallel to events.jsonl — and Tao no longer needs a separate TOY_DESCRIBE realize-only pre-pass. One graph walk, once. Backend-agnostic: ToyDescribeFlow reads the ggml graph structure via CPU TinyNN (same seam the CUDA/Metal checkpoint writer uses).



184
185
186
187
188
189
190
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 184

def self.emit_flow_json(run_dir, sess)
  if run_dir.length > 0
    File.open(run_dir + "/flow.json", "w") do |ff|
      ff.write(ToyDescribeFlow.json(sess))
    end
  end
end

.emit_one(t, is_node) ⇒ Object

Pure-string emit for one tensor.



154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 154

def self.emit_one(t, is_node)
  name  = TinyNN.tnn_tensor_name(t)
  flags = TinyNN.tnn_tensor_flags(t)
  dtype = TinyNN.tnn_tensor_dtype(t)
  nel   = TinyNN.tnn_tensor_nelements(t)
  nb    = TinyNN.tnn_tensor_nbytes(t)
  op    = TinyNN.tnn_tensor_op_name(t)
  role  = flag_role(flags, is_node)
  shape = shape_string(t)

  j  = "{"
  j = j + "\"name\":\""  + name + "\""
  j = j + ",\"role\":\"" + role + "\""
  j = j + ",\"shape\":\""+ shape + "\""
  j = j + ",\"dtype\":\""+ dtype_name(dtype) + "\""
  j = j + ",\"nelements\":" + nel.to_s
  j = j + ",\"nbytes\":"    + nb.to_s
  if is_node
    j = j + ",\"op\":\""   + op + "\""
  end
  j = j + "}"
  j
end

.flag_role(flags, is_node) ⇒ Object

ggml_tensor.flags bitmask → role string. PARAM > LOSS > OUTPUT > INPUT, then COMPUTE/leaf distinguished by caller.



52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 52

def self.flag_role(flags, is_node)
  if (flags & 4) != 0
    "weight"           # PARAM
  elsif (flags & 8) != 0
    "loss"             # LOSS
  elsif (flags & 2) != 0
    "output"           # OUTPUT
  elsif (flags & 1) != 0
    "input"            # INPUT
  elsif is_node
    "ephemeral"
  else
    "constant"
  end
end

.json(sess) ⇒ Object 



192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 192

def self.json(sess)
  pair    = build_index(sess)
  ptrs    = pair[0]
  n_leafs = pair[1]

  out = "{\"format\":\"toy/v1\",\"nodes\":["
  i = 0
  while i < ptrs.length
    out = out + (i == 0 ? "" : ",") + emit_one(ptrs[i], i >= n_leafs)
    i = i + 1
  end
  out = out + "],\"summary\":{"
  out = out + "\"n_leafs\":"  + n_leafs.to_s
  out = out + ",\"n_nodes\":" + (ptrs.length - n_leafs).to_s
  out = out + "}}"
  out
end

.mermaid(sess) ⇒ Object 



253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 253

def self.mermaid(sess)
  pair    = build_index(sess)
  ptrs    = pair[0]
  n_leafs = pair[1]
  n_nodes = ptrs.length - n_leafs

  out = "flowchart LR\n"

  i = 0
  while i < ptrs.length
    t = ptrs[i]
    is_node = i >= n_leafs
    flags = TinyNN.tnn_tensor_flags(t)
    role  = flag_role(flags, is_node)
    name  = TinyNN.tnn_tensor_name(t)
    shape = shape_string(t)
    op    = is_node ? TinyNN.tnn_tensor_op_name(t) : ""
    label = name.length > 0 ? name : "(unnamed)"
    if is_node && op.length > 0
      label = label + " " + op
    end
    label = label + " [" + shape + "]"
    glyph = "[" + label + "]"                      # default: ephemeral / output
    if role == "weight"
      glyph = "[[" + label + "]]"                  # subroutine = weight
    elsif role == "input"
      glyph = "([" + label + "])"                  # stadium = input
    elsif role == "constant"
      glyph = "(" + label + ")"                    # round = constant
    end
    out = out + "  t" + i.to_s + glyph + "\n"
    i = i + 1
  end

  # Edges (src → dst) for each compute node.
  i = 0
  while i < n_nodes
    dst_idx = n_leafs + i
    dst = ptrs[dst_idx]
    si = 0
    while si < 10
      src = TinyNN.tnn_tensor_src(dst, si)
      if src == nil || src == TinyNN.tnn_null_ptr
        si = 10
      else
        src_idx = -1
        j = 0
        while j < ptrs.length
          if ptrs[j] == src
            src_idx = j
            j = ptrs.length
          else
            j = j + 1
          end
        end
        if src_idx >= 0
          out = out + "  t" + src_idx.to_s + " --> t" + dst_idx.to_s + "\n"
        end
        si = si + 1
      end
    end
    i = i + 1
  end
  out
end

.param_count(sess) ⇒ Object

P1 — count the PARAM tensors in the session graph. Used by the derivation smoke as a cross-check against ToyDriftGrad.params.



397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 397

def self.param_count(sess)
  n = 0
  pair    = build_index(sess)
  ptrs    = pair[0]
  n_leafs = pair[1]
  i = 0
  while i < n_leafs
    flags = TinyNN.tnn_tensor_flags(ptrs[i])
    if (flags & 4) != 0
      n = n + 1
    end
    i = i + 1
  end
  n
end

.shape_string(t) ⇒ Object

Shape as “ne0×ne1×ne2×ne3” trimming trailing 1s.



69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 69

def self.shape_string(t)
  n0 = TinyNN.tnn_tensor_ne0(t)
  n1 = TinyNN.tnn_tensor_ne1(t)
  n2 = TinyNN.tnn_tensor_ne2(t)
  n3 = TinyNN.tnn_tensor_ne3(t)
  s = n0.to_s
  if n1 > 1 || n2 > 1 || n3 > 1
    s = s + "×" + n1.to_s
  end
  if n2 > 1 || n3 > 1
    s = s + "×" + n2.to_s
  end
  if n3 > 1
    s = s + "×" + n3.to_s
  end
  s
end

.text(sess) ⇒ Object 



210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
 
# File 'lib/toy/dev/toy_describe_flow.rb', line 210

def self.text(sess)
  pair    = build_index(sess)
  ptrs    = pair[0]
  n_leafs = pair[1]
  n_nodes = ptrs.length - n_leafs

  out = ""
  out = out + "DAG over session graph (toy/v1)\n"
  out = out + "  leafs: " + n_leafs.to_s + "   nodes: " + n_nodes.to_s + "\n"
  out = out + "\n"
  out = out + "LEAFS (weights / inputs / constants):\n"
  out = out + "  " + "role".ljust(10) + "dtype".ljust(7) +
              "shape".ljust(20) + "name\n"
  i = 0
  while i < n_leafs
    t = ptrs[i]
    flags = TinyNN.tnn_tensor_flags(t)
    role  = flag_role(flags, false)
    out = out + "  " + role.ljust(10) +
          dtype_name(TinyNN.tnn_tensor_dtype(t)).ljust(7) +
          shape_string(t).ljust(20) +
          TinyNN.tnn_tensor_name(t) + "\n"
    i = i + 1
  end
  out = out + "\n"
  out = out + "NODES (compute ops, in dispatch order):\n"
  out = out + "  " + "idx".ljust(5) + "role".ljust(10) + "dtype".ljust(7) +
              "shape".ljust(20) + "op".ljust(14) + "name\n"
  i = 0
  while i < n_nodes
    t = ptrs[n_leafs + i]
    flags = TinyNN.tnn_tensor_flags(t)
    role  = flag_role(flags, true)
    out = out + "  " + i.to_s.ljust(5) + role.ljust(10) +
          dtype_name(TinyNN.tnn_tensor_dtype(t)).ljust(7) +
          shape_string(t).ljust(20) +
          TinyNN.tnn_tensor_op_name(t).ljust(14) +
          TinyNN.tnn_tensor_name(t) + "\n"
    i = i + 1
  end
  out
end