codemaivanngu commited on
Commit
a1a9de3
·
1 Parent(s): e1515b2

fix(h1b): exclude non-EOS specials in offline row_groups

Browse files

Sampled reserved specials passed grouping and crashed exact_strprob in VocabIndex.bytes with KeyError. row_groups now marks such responses invalid with reason unsupported_special, using _special_token_ids exactly like the training_step post-filter. Adds T9b regression test; updates H1B_SOURCE hashes with fix entry. Numerical kernels and frozen Strict files unchanged.

Files changed (3) hide show
  1. docs/H1B_SOURCE.json +10 -2
  2. tests/test_h1b.py +19 -0
  3. tools/h1b/common.py +20 -2
docs/H1B_SOURCE.json CHANGED
@@ -52,7 +52,7 @@
52
  },
53
  {
54
  "path": "tools/h1b/common.py",
55
- "sha256": "8adf680d9e747e40f2781613a2b53b77f44b58bfe58ac0606c08432de4e05d06"
56
  },
57
  {
58
  "path": "tools/h1b/enrich_archive.py",
@@ -88,7 +88,7 @@
88
  },
89
  {
90
  "path": "tests/test_h1b.py",
91
- "sha256": "61e61fff4bdf4a8b1dea1c6e9fd97974a34d75b06a1516ce562dfe971b9a2c40"
92
  }
93
  ],
94
  "frozen_strict": [
@@ -125,5 +125,13 @@
125
  "scope": "Automatic runtime selection/ensure only; H1b numerical algorithm and kernels unchanged",
126
  "baseline_commit": "2bf0f47109fb9435232d77c1509eba0d894910ea",
127
  "path": "tools/h1b/canary.py"
 
 
 
 
 
 
 
 
128
  }
129
  }
 
52
  },
53
  {
54
  "path": "tools/h1b/common.py",
55
+ "sha256": "42947a601031445a49d63915ba3e5849cc8172d62fdec256fe9065e0f3596176"
56
  },
57
  {
58
  "path": "tools/h1b/enrich_archive.py",
 
88
  },
89
  {
90
  "path": "tests/test_h1b.py",
91
+ "sha256": "3e9bc8203c5f4fce2ea1bab21815bc184fbd8e972dc6d0b991dfbc6771710ed9"
92
  }
93
  ],
94
  "frozen_strict": [
 
125
  "scope": "Automatic runtime selection/ensure only; H1b numerical algorithm and kernels unchanged",
126
  "baseline_commit": "2bf0f47109fb9435232d77c1509eba0d894910ea",
127
  "path": "tools/h1b/canary.py"
128
+ },
129
+ "unsupported_special_offline_fix": {
130
+ "updated_sha256": {
131
+ "tools/h1b/common.py": "42947a601031445a49d63915ba3e5849cc8172d62fdec256fe9065e0f3596176",
132
+ "tests/test_h1b.py": "3e9bc8203c5f4fce2ea1bab21815bc184fbd8e972dc6d0b991dfbc6771710ed9"
133
+ },
134
+ "date": "2026-10-20",
135
+ "scope": "Offline row_groups now marks responses containing non-EOS special IDs invalid with reason unsupported_special, using _special_token_ids exactly like the training_step post-filter; previously such responses (e.g. sampled <|reserved_special_token_209|>) crashed exact_strprob in VocabIndex.bytes with KeyError. Numerical kernels and frozen Strict files unchanged"
136
  }
137
  }
tests/test_h1b.py CHANGED
@@ -58,6 +58,25 @@ class H1bTests(unittest.TestCase):
58
  a=([g.t_idx[0] for g in groups],[g.s_idx[0] for g in groups])
59
  self.assertEqual(a,align_byte_span_batch_1to1(tea,stu,tok.eos_token,tok.eos_token,codec,codec))
60
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
61
  def test_T4_trunk_and_exact_enumeration_with_duplicate_byte_ids(self):
62
  from kdflow.h1b.vocab_index import VocabIndex
63
  from kdflow.h1b.strprob import chunk_logprobs
 
58
  a=([g.t_idx[0] for g in groups],[g.s_idx[0] for g in groups])
59
  self.assertEqual(a,align_byte_span_batch_1to1(tea,stu,tok.eos_token,tok.eos_token,codec,codec))
60
 
61
+ def test_T9b_row_groups_rejects_unsupported_special_whole_response(self):
62
+ import torch
63
+ from tools.h1b.common import row_groups
64
+ # A sampled reserved special must invalidate the whole response with
65
+ # reason 'unsupported_special', matching the training_step post-filter,
66
+ # instead of reaching the byte vocabulary index (KeyError there).
67
+ class Tok(fixtures.Tokenizer):
68
+ pieces={0:'<eos>',1:'a',2:'b',7:'<pad>',8:'<rsvd>'}
69
+ all_special_ids=[0,7,8]
70
+ tok=Tok()
71
+ ids=torch.tensor([[1,2,0],[1,8,0]])
72
+ mask=torch.tensor([[True,True,False],[True,True,False]])
73
+ batch={'stu_input_ids':ids,'stu_loss_mask':mask,'stu_attn_mask':torch.ones_like(ids),
74
+ 'tea_input_ids':ids.clone(),'tea_loss_mask':mask.clone(),'tea_attn_mask':torch.ones_like(ids)}
75
+ rows=row_groups(batch,tok,tok)
76
+ self.assertTrue(rows[0].valid)
77
+ self.assertFalse(rows[1].valid)
78
+ self.assertEqual(rows[1].reason,'unsupported_special')
79
+
80
  def test_T4_trunk_and_exact_enumeration_with_duplicate_byte_ids(self):
81
  from kdflow.h1b.vocab_index import VocabIndex
82
  from kdflow.h1b.strprob import chunk_logprobs
tools/h1b/common.py CHANGED
@@ -67,14 +67,32 @@ def tokenize_record(record,student_tok,teacher_tok,assume_shared_prompt=False):
67
 
68
 
69
  def row_groups(batch,student_tok,teacher_tok):
70
- from kdflow.token_alignment import get_tokenizer_byte_codec
 
71
  from kdflow.h1b.groups import align_byte_span_groups_batch
72
  rows=[]
73
  for key,tok in (('tea',teacher_tok),('stu',student_tok)):
74
  labels=batch[key+'_input_ids'].roll(-1,1)
75
  rows.append([tok.convert_ids_to_tokens(labels[i][batch[key+'_loss_mask'][i].bool()].tolist()) for i in range(len(labels))])
76
- return align_byte_span_groups_batch(*rows,teacher_tok.eos_token,student_tok.eos_token,
77
  get_tokenizer_byte_codec(teacher_tok).piece_to_bytes,get_tokenizer_byte_codec(student_tok).piece_to_bytes)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
78
 
79
 
80
  def forward_hidden(model,batch,prefix):
 
67
 
68
 
69
  def row_groups(batch,student_tok,teacher_tok):
70
+ from dataclasses import replace
71
+ from kdflow.token_alignment import get_tokenizer_byte_codec,_special_token_ids
72
  from kdflow.h1b.groups import align_byte_span_groups_batch
73
  rows=[]
74
  for key,tok in (('tea',teacher_tok),('stu',student_tok)):
75
  labels=batch[key+'_input_ids'].roll(-1,1)
76
  rows.append([tok.convert_ids_to_tokens(labels[i][batch[key+'_loss_mask'][i].bool()].tolist()) for i in range(len(labels))])
77
+ grouped=align_byte_span_groups_batch(*rows,teacher_tok.eos_token,student_tok.eos_token,
78
  get_tokenizer_byte_codec(teacher_tok).piece_to_bytes,get_tokenizer_byte_codec(student_tok).piece_to_bytes)
79
+ # Twin of the H1bSpanCTKD.training_step post-filter: any non-EOS special ID
80
+ # invalidates the whole response, because the byte vocabulary index has no
81
+ # entry for specials (e.g. sampled <|reserved_special_token_*|>). Both sites
82
+ # must use _special_token_ids so the exclusion set can never drift apart.
83
+ tea_special,stu_special=_special_token_ids(teacher_tok),_special_token_ids(student_tok)
84
+ tea_eos,stu_eos=teacher_tok.eos_token_id,student_tok.eos_token_id
85
+ tea_ids=batch['tea_input_ids'].roll(-1,1); stu_ids=batch['stu_input_ids'].roll(-1,1)
86
+ tea_mask=batch['tea_loss_mask'].bool(); stu_mask=batch['stu_loss_mask'].bool()
87
+ fixed=[]
88
+ for i,row in enumerate(grouped):
89
+ tids=tea_ids[i][tea_mask[i]].tolist(); sids=stu_ids[i][stu_mask[i]].tolist()
90
+ if (any(t in tea_special and t!=tea_eos for t in tids)
91
+ or any(s in stu_special and s!=stu_eos for s in sids)):
92
+ fixed.append(replace(row,groups=(),valid=False,reason='unsupported_special'))
93
+ else:
94
+ fixed.append(row)
95
+ return fixed
96
 
97
 
98
  def forward_hidden(model,batch,prefix):