Add files using upload-large-folder tool

LTA_openwebtext_dualt/logs/eval_lm1b_latest_non_owt_methods_genppl_20260506_20260506_101041.log

@@ -0,0 +1,90 @@
+[load] mdlm_latest step=860000 ckpt=runs/lm1b_mdlm_unified_ddit_small_len128_gbs512_8gpu_1m_20260505_repro/latest.pt
+[mdlm] generated 32/256
+[mdlm] generated 64/256
+[mdlm] generated 96/256
+[mdlm] generated 128/256
+[mdlm] generated 160/256
+[mdlm] generated 192/256
+[mdlm] generated 224/256
+[mdlm] generated 256/256
+[summary] {"type": "summary", "name": "mdlm_latest", "kind": "mdlm", "checkpoint": "runs/lm1b_mdlm_unified_ddit_small_len128_gbs512_8gpu_1m_20260505_repro/latest.pt", "step": 860000, "decode": {"kind": "mdlm", "steps": 256, "decode_rule": "stochastic_confidence_unmask", "start": "all_mask", "temp": 1.0, "selection_noise": 0.0, "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 13.254623394640433, "nll_per_token": 2.584346427012978, "tokens": 38421, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 15.945599491226918, "nll_per_token": 2.769182897198147, "tokens": 31758, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 3.5550877572540878, "unique_tokens": 2672, "token_count": 32768, "distinct_1": 0.08154296875, "distinct_2": 0.27294537401574803, "top_token_mass": 0.182861328125}}
+[load] duo_latest step=698000 ckpt=runs/lm1b_duo_unified_ddit_small_len128_gbs512_8gpu_1m_20260505_repro/latest.pt
+[duo discrete] generated 32/256
+[duo discrete] generated 64/256
+[duo discrete] generated 96/256
+[duo discrete] generated 128/256
+[duo discrete] generated 160/256
+[duo discrete] generated 192/256
+[duo discrete] generated 224/256
+[duo discrete] generated 256/256
+[summary] {"type": "summary", "name": "duo_latest", "kind": "duo", "checkpoint": "runs/lm1b_duo_unified_ddit_small_len128_gbs512_8gpu_1m_20260505_repro/latest.pt", "step": 698000, "decode": {"kind": "duo", "steps": 256, "decode_rule": "discrete_random_replace_confidence_lock", "temp": 1.0, "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 1.10513806766837, "nll_per_token": 0.09997027528052237, "tokens": 65280, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 1.0, "nll_per_token": 0.0, "tokens": 0, "kept_samples": 0, "total_samples": 0, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 0.0, "unique_tokens": 1, "token_count": 32768, "distinct_1": 3.0517578125e-05, "distinct_2": 3.075787401574803e-05, "top_token_mass": 1.0}}
+[load] flm_latest step=369000 ckpt=runs/lm1b_flm_unified_ddit_small_len128_gbs512_8gpu_1m_20260506_repro/latest.pt
+[flm] generated 32/256
+[flm] generated 64/256
+[flm] generated 96/256
+[flm] generated 128/256
+[flm] generated 160/256
+[flm] generated 192/256
+[flm] generated 224/256
+[flm] generated 256/256
+[summary] {"type": "summary", "name": "flm_latest", "kind": "flm", "checkpoint": "runs/lm1b_flm_unified_ddit_small_len128_gbs512_8gpu_1m_20260506_repro/latest.pt", "step": 369000, "decode": {"kind": "flm", "steps": 256, "decode_rule": "gaussian_to_endpoint_flow", "noise_init": "gaussian", "final_from": "blend", "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 149.36539937152122, "nll_per_token": 5.006395648630426, "tokens": 38794, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 279.18684360473856, "nll_per_token": 5.63188124801842, "tokens": 32329, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 4.3123598709846584, "unique_tokens": 5896, "token_count": 32768, "distinct_1": 0.179931640625, "distinct_2": 0.6859005905511811, "top_token_mass": 0.045074462890625}}
+[load] categorical_fullvocab_latest step=609000 ckpt=runs/lta_lm1b_dirichlet_categorical_fullvocab_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/latest.pt
+[categorical Cmax=1024] generated 32/256
+[categorical Cmax=1024] generated 64/256
+[categorical Cmax=1024] generated 96/256
+[categorical Cmax=1024] generated 128/256
+[categorical Cmax=1024] generated 160/256
+[categorical Cmax=1024] generated 192/256
+[categorical Cmax=1024] generated 224/256
+[categorical Cmax=1024] generated 256/256
+[summary] {"type": "summary", "name": "categorical_fullvocab_latest", "kind": "categorical_fullvocab", "checkpoint": "runs/lta_lm1b_dirichlet_categorical_fullvocab_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/latest.pt", "step": 609000, "decode": {"kind": "categorical_fullvocab", "steps": 1024, "model_t_mode": "const05", "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.5, "anchor_mode": "state", "noise_init": "dirichlet", "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "endpoint_temp": 1.3, "final_from": "blend", "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 26.097635854173753, "nll_per_token": 3.2618447299261484, "tokens": 38904, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 37.64924659414174, "nll_per_token": 3.6283129432659162, "tokens": 31619, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 4.01496464321082, "unique_tokens": 2943, "token_count": 32768, "distinct_1": 0.089813232421875, "distinct_2": 0.421751968503937, "top_token_mass": 0.0771484375}}
+[load] categorical_fullvocab_c64_latest step=301000 ckpt=runs/lta_lm1b_dirichlet_categorical_fullvocab_c64p0_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/latest.pt
+[categorical Cmax=64] generated 32/256
+[categorical Cmax=64] generated 64/256
+[categorical Cmax=64] generated 96/256
+[categorical Cmax=64] generated 128/256
+[categorical Cmax=64] generated 160/256
+[categorical Cmax=64] generated 192/256
+[categorical Cmax=64] generated 224/256
+[categorical Cmax=64] generated 256/256
+[summary] {"type": "summary", "name": "categorical_fullvocab_c64_latest", "kind": "categorical_fullvocab", "checkpoint": "runs/lta_lm1b_dirichlet_categorical_fullvocab_c64p0_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/latest.pt", "step": 301000, "decode": {"kind": "categorical_fullvocab", "steps": 1024, "model_t_mode": "const05", "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.5, "anchor_mode": "state", "noise_init": "dirichlet", "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 64.0, "endpoint_temp": 1.3, "final_from": "blend", "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 29.50899600665449, "nll_per_token": 3.3846951662362814, "tokens": 38725, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 43.028414020337806, "nll_per_token": 3.761860688637853, "tokens": 31666, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 3.9699503322445215, "unique_tokens": 2842, "token_count": 32768, "distinct_1": 0.08673095703125, "distinct_2": 0.41627706692913385, "top_token_mass": 0.0784912109375}}
+[load] categorical_fullvocab_c256_latest step=264000 ckpt=runs/lta_lm1b_dirichlet_categorical_fullvocab_c256p0_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/latest.pt
+[categorical Cmax=256] generated 32/256
+[categorical Cmax=256] generated 64/256
+[categorical Cmax=256] generated 96/256
+[categorical Cmax=256] generated 128/256
+[categorical Cmax=256] generated 160/256
+[categorical Cmax=256] generated 192/256
+[categorical Cmax=256] generated 224/256
+[categorical Cmax=256] generated 256/256
+[summary] {"type": "summary", "name": "categorical_fullvocab_c256_latest", "kind": "categorical_fullvocab", "checkpoint": "runs/lta_lm1b_dirichlet_categorical_fullvocab_c256p0_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/latest.pt", "step": 264000, "decode": {"kind": "categorical_fullvocab", "steps": 1024, "model_t_mode": "const05", "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.5, "anchor_mode": "state", "noise_init": "dirichlet", "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 256.0, "endpoint_temp": 1.3, "final_from": "blend", "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 29.0563917263388, "nll_per_token": 3.369238484099026, "tokens": 39236, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 42.82775902266668, "nll_per_token": 3.757186467649641, "tokens": 31964, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 3.9620396608009667, "unique_tokens": 3100, "token_count": 32768, "distinct_1": 0.0946044921875, "distinct_2": 0.42165969488188976, "top_token_mass": 0.091522216796875}}
+[load] categorical_fullvocab_c16_4gpu_latest step=146000 ckpt=runs/lta_lm1b_dirichlet_categorical_fullvocab_c16p0_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_4gpu_1m_nw0/latest.pt
+[categorical Cmax=16] generated 32/256
+[categorical Cmax=16] generated 64/256
+[categorical Cmax=16] generated 96/256
+[categorical Cmax=16] generated 128/256
+[categorical Cmax=16] generated 160/256
+[categorical Cmax=16] generated 192/256
+[categorical Cmax=16] generated 224/256
+[categorical Cmax=16] generated 256/256
+[summary] {"type": "summary", "name": "categorical_fullvocab_c16_4gpu_latest", "kind": "categorical_fullvocab", "checkpoint": "runs/lta_lm1b_dirichlet_categorical_fullvocab_c16p0_dualt_flmpack_onehot_hardce_ddit_small_len128_gbs512_4gpu_1m_nw0/latest.pt", "step": 146000, "decode": {"kind": "categorical_fullvocab", "steps": 1024, "model_t_mode": "const05", "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.5, "anchor_mode": "state", "noise_init": "dirichlet", "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 16.0, "endpoint_temp": 1.3, "final_from": "blend", "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 34.32179659456556, "nll_per_token": 3.535780621576138, "tokens": 38197, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 50.163586453004264, "nll_per_token": 3.915289394027899, "tokens": 31430, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 3.954715470318628, "unique_tokens": 2829, "token_count": 32768, "distinct_1": 0.086334228515625, "distinct_2": 0.4382689468503937, "top_token_mass": 0.07720947265625}}
+[load] ar_rowshard_latest step=267000 ckpt=runs/ar_lm1b_flmpack_bert_small_len128_gbs512_4gpu_1m_rowshard_b64_resume4000_20260504_203021/latest.pt
+[ar temp=0.8] generated 32/256
+[ar temp=0.8] generated 64/256
+[ar temp=0.8] generated 96/256
+[ar temp=0.8] generated 128/256
+[ar temp=0.8] generated 160/256
+[ar temp=0.8] generated 192/256
+[ar temp=0.8] generated 224/256
+[ar temp=0.8] generated 256/256
+[ar temp=1] generated 32/256
+[ar temp=1] generated 64/256
+[ar temp=1] generated 96/256
+[ar temp=1] generated 128/256
+[ar temp=1] generated 160/256
+[ar temp=1] generated 192/256
+[ar temp=1] generated 224/256
+[ar temp=1] generated 256/256
+[summary] {"type": "summary", "name": "ar_rowshard_latest_t0p8", "kind": "ar", "checkpoint": "runs/ar_lm1b_flmpack_bert_small_len128_gbs512_4gpu_1m_rowshard_b64_resume4000_20260504_203021/latest.pt", "step": 267000, "decode": {"kind": "ar_sample", "temp": 0.8, "max_new_tokens": 127, "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 32.75331929719446, "nll_per_token": 3.489004309531525, "tokens": 38672, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 46.344570351610095, "nll_per_token": 3.8361041406323277, "tokens": 31984, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 4.1883945709687715, "unique_tokens": 5282, "token_count": 32768, "distinct_1": 0.16119384765625, "distinct_2": 0.590089812992126, "top_token_mass": 0.065338134765625}}
+[summary] {"type": "summary", "name": "ar_rowshard_latest_t1p0", "kind": "ar", "checkpoint": "runs/ar_lm1b_flmpack_bert_small_len128_gbs512_4gpu_1m_rowshard_b64_resume4000_20260504_203021/latest.pt", "step": 267000, "decode": {"kind": "ar_sample", "temp": 1.0, "max_new_tokens": 127, "n_samples": 256, "seed": 20260506}, "raw_genppl": {"ppl": 70.47566606515413, "nll_per_token": 4.255267488039646, "tokens": 38787, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 111.28221270824383, "nll_per_token": 4.7120694315677865, "tokens": 32392, "kept_samples": 256, "total_samples": 256, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 4.344022305264332, "unique_tokens": 6845, "token_count": 32768, "distinct_1": 0.208892822265625, "distinct_2": 0.7198572834645669, "top_token_mass": 0.043731689453125}}
+[done] docs/lta_samples/metrics_20260506/latest_non_owt_all_methods

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/pygments/lexers/ada.py

@@ -0,0 +1,144 @@
+"""
+    pygments.lexers.ada
+    ~~~~~~~~~~~~~~~~~~~
+
+    Lexers for Ada family languages.
+
+    :copyright: Copyright 2006-present by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+import re
+
+from pygments.lexer import RegexLexer, include, bygroups, words, using, this, \
+    default
+from pygments.token import Text, Comment, Operator, Keyword, Name, String, \
+    Number, Punctuation
+from pygments.lexers._ada_builtins import KEYWORD_LIST, BUILTIN_LIST
+
+__all__ = ['AdaLexer']
+
+
+class AdaLexer(RegexLexer):
+    """
+    For Ada source code.
+    """
+
+    name = 'Ada'
+    aliases = ['ada', 'ada95', 'ada2005']
+    filenames = ['*.adb', '*.ads', '*.ada']
+    mimetypes = ['text/x-ada']
+    url = 'https://www.adaic.org'
+    version_added = '1.3'
+
+    flags = re.MULTILINE | re.IGNORECASE
+
+    tokens = {
+        'root': [
+            (r'[^\S\n]+', Text),
+            (r'--.*?\n', Comment.Single),
+            (r'[^\S\n]+', Text),
+            (r'function|procedure|entry', Keyword.Declaration, 'subprogram'),
+            (r'(subtype|type)(\s+)(\w+)',
+             bygroups(Keyword.Declaration, Text, Keyword.Type), 'type_def'),
+            (r'task|protected', Keyword.Declaration),
+            (r'(subtype)(\s+)', bygroups(Keyword.Declaration, Text)),
+            (r'(end)(\s+)', bygroups(Keyword.Reserved, Text), 'end'),
+            (r'(pragma)(\s+)(\w+)', bygroups(Keyword.Reserved, Text,
+                                             Comment.Preproc)),
+            (r'(true|false|null)\b', Keyword.Constant),
+            # builtin types
+            (words(BUILTIN_LIST, suffix=r'\b'), Keyword.Type),
+            (r'(and(\s+then)?|in|mod|not|or(\s+else)|rem)\b', Operator.Word),
+            (r'generic|private', Keyword.Declaration),
+            (r'package', Keyword.Declaration, 'package'),
+            (r'array\b', Keyword.Reserved, 'array_def'),
+            (r'(with|use)(\s+)', bygroups(Keyword.Namespace, Text), 'import'),
+            (r'(\w+)(\s*)(:)(\s*)(constant)',
+             bygroups(Name.Constant, Text, Punctuation, Text,
+                      Keyword.Reserved)),
+            (r'<<\w+>>', Name.Label),
+            (r'(\w+)(\s*)(:)(\s*)(declare|begin|loop|for|while)',
+             bygroups(Name.Label, Text, Punctuation, Text, Keyword.Reserved)),
+            # keywords
+            (words(KEYWORD_LIST, prefix=r'\b', suffix=r'\b'),
+             Keyword.Reserved),
+            (r'"[^"]*"', String),
+            include('attribute'),
+            include('numbers'),
+            (r"'[^']'", String.Character),
+            (r'(\w+)(\s*|[(,])', bygroups(Name, using(this))),
+            (r"(<>|=>|:=|@|[\[\]]|[()|:;,.'])", Punctuation),
+            (r'[*<>+=/&-]', Operator),
+            (r'\n+', Text),
+        ],
+        'numbers': [
+            (r'[0-9_]+#[0-9a-f_\.]+#', Number.Hex),
+            (r'[0-9_]+\.[0-9_]*', Number.Float),
+            (r'[0-9_]+', Number.Integer),
+        ],
+        'attribute': [
+            (r"(')(\w+)", bygroups(Punctuation, Name.Attribute)),
+        ],
+        'subprogram': [
+            (r'\(', Punctuation, ('#pop', 'formal_part')),
+            (r';', Punctuation, '#pop'),
+            (r'is\b', Keyword.Reserved, '#pop'),
+            (r'"[^"]+"|\w+', Name.Function),
+            include('root'),
+        ],
+        'end': [
+            ('(if|case|record|loop|select)', Keyword.Reserved),
+            (r'"[^"]+"|[\w.]+', Name.Function),
+            (r'\s+', Text),
+            (';', Punctuation, '#pop'),
+        ],
+        'type_def': [
+            (r';', Punctuation, '#pop'),
+            (r'\(', Punctuation, 'formal_part'),
+            (r'\[', Punctuation, 'formal_part'),
+            (r'with|and|use', Keyword.Reserved),
+            (r'array\b', Keyword.Reserved, ('#pop', 'array_def')),
+            (r'record\b', Keyword.Reserved, ('record_def')),
+            (r'(null record)(;)', bygroups(Keyword.Reserved, Punctuation), '#pop'),
+            include('root'),
+        ],
+        'array_def': [
+            (r';', Punctuation, '#pop'),
+            (r'(\w+)(\s+)(range)', bygroups(Keyword.Type, Text, Keyword.Reserved)),
+            include('root'),
+        ],
+        'record_def': [
+            (r'end record', Keyword.Reserved, '#pop'),
+            include('root'),
+        ],
+        'import': [
+            # TODO: use Name.Namespace if appropriate.  This needs
+            # work to disinguish imports from aspects.
+            (r'[\w.]+', Name, '#pop'),
+            default('#pop'),
+        ],
+        'formal_part': [
+            (r'\)', Punctuation, '#pop'),
+            (r'\]', Punctuation, '#pop'),
+            (r'\w+', Name.Variable),
+            (r',|:[^=]', Punctuation),
+            (r'(in|not|null|out|access)\b', Keyword.Reserved),
+            include('root'),
+        ],
+        'package': [
+            ('body', Keyword.Declaration),
+            (r'is\s+new|renames', Keyword.Reserved),
+            ('is', Keyword.Reserved, '#pop'),
+            (';', Punctuation, '#pop'),
+            (r'\(', Punctuation, 'package_instantiation'),
+            (r'([\w.]+)', Name.Class),
+            include('root'),
+        ],
+        'package_instantiation': [
+            (r'("[^"]+"|\w+)(\s+)(=>)', bygroups(Name.Variable, Text, Punctuation)),
+            (r'[\w.\'"]', Text),
+            (r'\)', Punctuation, '#pop'),
+            include('root'),
+        ],
+    }

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/pygments/lexers/elm.py

@@ -0,0 +1,123 @@
+"""
+    pygments.lexers.elm
+    ~~~~~~~~~~~~~~~~~~~
+
+    Lexer for the Elm programming language.
+
+    :copyright: Copyright 2006-present by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+from pygments.lexer import RegexLexer, words, include, bygroups
+from pygments.token import Comment, Keyword, Name, Number, Punctuation, \
+    String, Whitespace
+
+__all__ = ['ElmLexer']
+
+
+class ElmLexer(RegexLexer):
+    """
+    For Elm source code.
+    """
+
+    name = 'Elm'
+    url = 'https://elm-lang.org/'
+    aliases = ['elm']
+    filenames = ['*.elm']
+    mimetypes = ['text/x-elm']
+    version_added = '2.1'
+
+    validName = r'[a-z_][a-zA-Z0-9_\']*'
+
+    specialName = r'^main '
+
+    builtinOps = (
+        '~', '||', '|>', '|', '`', '^', '\\', '\'', '>>', '>=', '>', '==',
+        '=', '<~', '<|', '<=', '<<', '<-', '<', '::', ':', '/=', '//', '/',
+        '..', '.', '->', '-', '++', '+', '*', '&&', '%',
+    )
+
+    reservedWords = words((
+        'alias', 'as', 'case', 'else', 'if', 'import', 'in',
+        'let', 'module', 'of', 'port', 'then', 'type', 'where',
+    ), suffix=r'\b')
+
+    tokens = {
+        'root': [
+
+            # Comments
+            (r'\{-', Comment.Multiline, 'comment'),
+            (r'--.*', Comment.Single),
+
+            # Whitespace
+            (r'\s+', Whitespace),
+
+            # Strings
+            (r'"', String, 'doublequote'),
+
+            # Modules
+            (r'^(\s*)(module)(\s*)', bygroups(Whitespace, Keyword.Namespace,
+                Whitespace), 'imports'),
+
+            # Imports
+            (r'^(\s*)(import)(\s*)', bygroups(Whitespace, Keyword.Namespace,
+                Whitespace), 'imports'),
+
+            # Shaders
+            (r'\[glsl\|.*', Name.Entity, 'shader'),
+
+            # Keywords
+            (reservedWords, Keyword.Reserved),
+
+            # Types
+            (r'[A-Z][a-zA-Z0-9_]*', Keyword.Type),
+
+            # Main
+            (specialName, Keyword.Reserved),
+
+            # Prefix Operators
+            (words((builtinOps), prefix=r'\(', suffix=r'\)'), Name.Function),
+
+            # Infix Operators
+            (words(builtinOps), Name.Function),
+
+            # Numbers
+            include('numbers'),
+
+            # Variable Names
+            (validName, Name.Variable),
+
+            # Parens
+            (r'[,()\[\]{}]', Punctuation),
+
+        ],
+
+        'comment': [
+            (r'-(?!\})', Comment.Multiline),
+            (r'\{-', Comment.Multiline, 'comment'),
+            (r'[^-}]', Comment.Multiline),
+            (r'-\}', Comment.Multiline, '#pop'),
+        ],
+
+        'doublequote': [
+            (r'\\u[0-9a-fA-F]{4}', String.Escape),
+            (r'\\[nrfvb\\"]', String.Escape),
+            (r'[^"]', String),
+            (r'"', String, '#pop'),
+        ],
+
+        'imports': [
+            (r'\w+(\.\w+)*', Name.Class, '#pop'),
+        ],
+
+        'numbers': [
+            (r'_?\d+\.(?=\d+)', Number.Float),
+            (r'_?\d+', Number.Integer),
+        ],
+
+        'shader': [
+            (r'\|(?!\])', Name.Entity),
+            (r'\|\]', Name.Entity, '#pop'),
+            (r'(.*)(\n)', bygroups(Name.Entity, Whitespace)),
+        ],
+    }

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/pygments/lexers/modula2.py

@@ -0,0 +1,1579 @@
+"""
+    pygments.lexers.modula2
+    ~~~~~~~~~~~~~~~~~~~~~~~
+
+    Multi-Dialect Lexer for Modula-2.
+
+    :copyright: Copyright 2006-present by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+import re
+
+from pygments.lexer import RegexLexer, include
+from pygments.util import get_bool_opt, get_list_opt
+from pygments.token import Text, Comment, Operator, Keyword, Name, \
+    String, Number, Punctuation, Error
+
+__all__ = ['Modula2Lexer']
+
+
+# Multi-Dialect Modula-2 Lexer
+class Modula2Lexer(RegexLexer):
+    """
+    For Modula-2 source code.
+
+    The Modula-2 lexer supports several dialects.  By default, it operates in
+    fallback mode, recognising the *combined* literals, punctuation symbols
+    and operators of all supported dialects, and the *combined* reserved words
+    and builtins of PIM Modula-2, ISO Modula-2 and Modula-2 R10, while not
+    differentiating between library defined identifiers.
+
+    To select a specific dialect, a dialect option may be passed
+    or a dialect tag may be embedded into a source file.
+
+    Dialect Options:
+
+    `m2pim`
+        Select PIM Modula-2 dialect.
+    `m2iso`
+        Select ISO Modula-2 dialect.
+    `m2r10`
+        Select Modula-2 R10 dialect.
+    `objm2`
+        Select Objective Modula-2 dialect.
+
+    The PIM and ISO dialect options may be qualified with a language extension.
+
+    Language Extensions:
+
+    `+aglet`
+        Select Aglet Modula-2 extensions, available with m2iso.
+    `+gm2`
+        Select GNU Modula-2 extensions, available with m2pim.
+    `+p1`
+        Select p1 Modula-2 extensions, available with m2iso.
+    `+xds`
+        Select XDS Modula-2 extensions, available with m2iso.
+
+
+    Passing a Dialect Option via Unix Commandline Interface
+
+    Dialect options may be passed to the lexer using the `dialect` key.
+    Only one such option should be passed. If multiple dialect options are
+    passed, the first valid option is used, any subsequent options are ignored.
+
+    Examples:
+
+    `$ pygmentize -O full,dialect=m2iso -f html -o /path/to/output /path/to/input`
+        Use ISO dialect to render input to HTML output
+    `$ pygmentize -O full,dialect=m2iso+p1 -f rtf -o /path/to/output /path/to/input`
+        Use ISO dialect with p1 extensions to render input to RTF output
+
+
+    Embedding a Dialect Option within a source file
+
+    A dialect option may be embedded in a source file in form of a dialect
+    tag, a specially formatted comment that specifies a dialect option.
+
+    Dialect Tag EBNF::
+
+       dialectTag :
+           OpeningCommentDelim Prefix dialectOption ClosingCommentDelim ;
+
+       dialectOption :
+           'm2pim' | 'm2iso' | 'm2r10' | 'objm2' |
+           'm2iso+aglet' | 'm2pim+gm2' | 'm2iso+p1' | 'm2iso+xds' ;
+
+       Prefix : '!' ;
+
+       OpeningCommentDelim : '(*' ;
+
+       ClosingCommentDelim : '*)' ;
+
+    No whitespace is permitted between the tokens of a dialect tag.
+
+    In the event that a source file contains multiple dialect tags, the first
+    tag that contains a valid dialect option will be used and any subsequent
+    dialect tags will be ignored.  Ideally, a dialect tag should be placed
+    at the beginning of a source file.
+
+    An embedded dialect tag overrides a dialect option set via command line.
+
+    Examples:
+
+    ``(*!m2r10*) DEFINITION MODULE Foobar; ...``
+        Use Modula2 R10 dialect to render this source file.
+    ``(*!m2pim+gm2*) DEFINITION MODULE Bazbam; ...``
+        Use PIM dialect with GNU extensions to render this source file.
+
+
+    Algol Publication Mode:
+
+    In Algol publication mode, source text is rendered for publication of
+    algorithms in scientific papers and academic texts, following the format
+    of the Revised Algol-60 Language Report.  It is activated by passing
+    one of two corresponding styles as an option:
+
+    `algol`
+        render reserved words lowercase underline boldface
+        and builtins lowercase boldface italic
+    `algol_nu`
+        render reserved words lowercase boldface (no underlining)
+        and builtins lowercase boldface italic
+
+    The lexer automatically performs the required lowercase conversion when
+    this mode is activated.
+
+    Example:
+
+    ``$ pygmentize -O full,style=algol -f latex -o /path/to/output /path/to/input``
+        Render input file in Algol publication mode to LaTeX output.
+
+
+    Rendering Mode of First Class ADT Identifiers:
+
+    The rendering of standard library first class ADT identifiers is controlled
+    by option flag "treat_stdlib_adts_as_builtins".
+
+    When this option is turned on, standard library ADT identifiers are rendered
+    as builtins.  When it is turned off, they are rendered as ordinary library
+    identifiers.
+
+    `treat_stdlib_adts_as_builtins` (default: On)
+
+    The option is useful for dialects that support ADTs as first class objects
+    and provide ADTs in the standard library that would otherwise be built-in.
+
+    At present, only Modula-2 R10 supports library ADTs as first class objects
+    and therefore, no ADT identifiers are defined for any other dialects.
+
+    Example:
+
+    ``$ pygmentize -O full,dialect=m2r10,treat_stdlib_adts_as_builtins=Off ...``
+        Render standard library ADTs as ordinary library types.
+
+    .. versionchanged:: 2.1
+       Added multi-dialect support.
+    """
+    name = 'Modula-2'
+    url = 'http://www.modula2.org/'
+    aliases = ['modula2', 'm2']
+    filenames = ['*.def', '*.mod']
+    mimetypes = ['text/x-modula2']
+    version_added = '1.3'
+
+    flags = re.MULTILINE | re.DOTALL
+
+    tokens = {
+        'whitespace': [
+            (r'\n+', Text),  # blank lines
+            (r'\s+', Text),  # whitespace
+        ],
+        'dialecttags': [
+            # PIM Dialect Tag
+            (r'\(\*!m2pim\*\)', Comment.Special),
+            # ISO Dialect Tag
+            (r'\(\*!m2iso\*\)', Comment.Special),
+            # M2R10 Dialect Tag
+            (r'\(\*!m2r10\*\)', Comment.Special),
+            # ObjM2 Dialect Tag
+            (r'\(\*!objm2\*\)', Comment.Special),
+            # Aglet Extensions Dialect Tag
+            (r'\(\*!m2iso\+aglet\*\)', Comment.Special),
+            # GNU Extensions Dialect Tag
+            (r'\(\*!m2pim\+gm2\*\)', Comment.Special),
+            # p1 Extensions Dialect Tag
+            (r'\(\*!m2iso\+p1\*\)', Comment.Special),
+            # XDS Extensions Dialect Tag
+            (r'\(\*!m2iso\+xds\*\)', Comment.Special),
+        ],
+        'identifiers': [
+            (r'([a-zA-Z_$][\w$]*)', Name),
+        ],
+        'prefixed_number_literals': [
+            #
+            # Base-2, whole number
+            (r'0b[01]+(\'[01]+)*', Number.Bin),
+            #
+            # Base-16, whole number
+            (r'0[ux][0-9A-F]+(\'[0-9A-F]+)*', Number.Hex),
+        ],
+        'plain_number_literals': [
+            #
+            # Base-10, real number with exponent
+            (r'[0-9]+(\'[0-9]+)*'  # integral part
+             r'\.[0-9]+(\'[0-9]+)*'  # fractional part
+             r'[eE][+-]?[0-9]+(\'[0-9]+)*',  # exponent
+             Number.Float),
+            #
+            # Base-10, real number without exponent
+            (r'[0-9]+(\'[0-9]+)*'  # integral part
+             r'\.[0-9]+(\'[0-9]+)*',  # fractional part
+             Number.Float),
+            #
+            # Base-10, whole number
+            (r'[0-9]+(\'[0-9]+)*', Number.Integer),
+        ],
+        'suffixed_number_literals': [
+            #
+            # Base-8, whole number
+            (r'[0-7]+B', Number.Oct),
+            #
+            # Base-8, character code
+            (r'[0-7]+C', Number.Oct),
+            #
+            # Base-16, number
+            (r'[0-9A-F]+H', Number.Hex),
+        ],
+        'string_literals': [
+            (r'"(\\\\|\\[^\\]|[^"\\])*"', String.Double),
+            (r"'(\\\\|\\[^\\]|[^'\\])*'", String.Single),
+        ],
+        'digraph_operators': [
+            # Dot Product Operator
+            (r'\*\.', Operator),
+            # Array Concatenation Operator
+            (r'\+>', Operator),  # M2R10 + ObjM2
+            # Inequality Operator
+            (r'<>', Operator),  # ISO + PIM
+            # Less-Or-Equal, Subset
+            (r'<=', Operator),
+            # Greater-Or-Equal, Superset
+            (r'>=', Operator),
+            # Identity Operator
+            (r'==', Operator),  # M2R10 + ObjM2
+            # Type Conversion Operator
+            (r'::', Operator),  # M2R10 + ObjM2
+            # Assignment Symbol
+            (r':=', Operator),
+            # Postfix Increment Mutator
+            (r'\+\+', Operator),  # M2R10 + ObjM2
+            # Postfix Decrement Mutator
+            (r'--', Operator),  # M2R10 + ObjM2
+        ],
+        'unigraph_operators': [
+            # Arithmetic Operators
+            (r'[+-]', Operator),
+            (r'[*/]', Operator),
+            # ISO 80000-2 compliant Set Difference Operator
+            (r'\\', Operator),  # M2R10 + ObjM2
+            # Relational Operators
+            (r'[=#<>]', Operator),
+            # Dereferencing Operator
+            (r'\^', Operator),
+            # Dereferencing Operator Synonym
+            (r'@', Operator),  # ISO
+            # Logical AND Operator Synonym
+            (r'&', Operator),  # PIM + ISO
+            # Logical NOT Operator Synonym
+            (r'~', Operator),  # PIM + ISO
+            # Smalltalk Message Prefix
+            (r'`', Operator),  # ObjM2
+        ],
+        'digraph_punctuation': [
+            # Range Constructor
+            (r'\.\.', Punctuation),
+            # Opening Chevron Bracket
+            (r'<<', Punctuation),  # M2R10 + ISO
+            # Closing Chevron Bracket
+            (r'>>', Punctuation),  # M2R10 + ISO
+            # Blueprint Punctuation
+            (r'->', Punctuation),  # M2R10 + ISO
+            # Distinguish |# and # in M2 R10
+            (r'\|#', Punctuation),
+            # Distinguish ## and # in M2 R10
+            (r'##', Punctuation),
+            # Distinguish |* and * in M2 R10
+            (r'\|\*', Punctuation),
+        ],
+        'unigraph_punctuation': [
+            # Common Punctuation
+            (r'[()\[\]{},.:;|]', Punctuation),
+            # Case Label Separator Synonym
+            (r'!', Punctuation),  # ISO
+            # Blueprint Punctuation
+            (r'\?', Punctuation),  # M2R10 + ObjM2
+        ],
+        'comments': [
+            # Single Line Comment
+            (r'^//.*?\n', Comment.Single),  # M2R10 + ObjM2
+            # Block Comment
+            (r'\(\*([^$].*?)\*\)', Comment.Multiline),
+            # Template Block Comment
+            (r'/\*(.*?)\*/', Comment.Multiline),  # M2R10 + ObjM2
+        ],
+        'pragmas': [
+            # ISO Style Pragmas
+            (r'<\*.*?\*>', Comment.Preproc),  # ISO, M2R10 + ObjM2
+            # Pascal Style Pragmas
+            (r'\(\*\$.*?\*\)', Comment.Preproc),  # PIM
+        ],
+        'root': [
+            include('whitespace'),
+            include('dialecttags'),
+            include('pragmas'),
+            include('comments'),
+            include('identifiers'),
+            include('suffixed_number_literals'),  # PIM + ISO
+            include('prefixed_number_literals'),  # M2R10 + ObjM2
+            include('plain_number_literals'),
+            include('string_literals'),
+            include('digraph_punctuation'),
+            include('digraph_operators'),
+            include('unigraph_punctuation'),
+            include('unigraph_operators'),
+        ]
+    }
+
+#  C o m m o n   D a t a s e t s
+
+    # Common Reserved Words Dataset
+    common_reserved_words = (
+        # 37 common reserved words
+        'AND', 'ARRAY', 'BEGIN', 'BY', 'CASE', 'CONST', 'DEFINITION', 'DIV',
+        'DO', 'ELSE', 'ELSIF', 'END', 'EXIT', 'FOR', 'FROM', 'IF',
+        'IMPLEMENTATION', 'IMPORT', 'IN', 'LOOP', 'MOD', 'MODULE', 'NOT',
+        'OF', 'OR', 'POINTER', 'PROCEDURE', 'RECORD', 'REPEAT', 'RETURN',
+        'SET', 'THEN', 'TO', 'TYPE', 'UNTIL', 'VAR', 'WHILE',
+    )
+
+    # Common Builtins Dataset
+    common_builtins = (
+        # 16 common builtins
+        'ABS', 'BOOLEAN', 'CARDINAL', 'CHAR', 'CHR', 'FALSE', 'INTEGER',
+        'LONGINT', 'LONGREAL', 'MAX', 'MIN', 'NIL', 'ODD', 'ORD', 'REAL',
+        'TRUE',
+    )
+
+    # Common Pseudo-Module Builtins Dataset
+    common_pseudo_builtins = (
+        # 4 common pseudo builtins
+        'ADDRESS', 'BYTE', 'WORD', 'ADR'
+    )
+
+#  P I M   M o d u l a - 2   D a t a s e t s
+
+    # Lexemes to Mark as Error Tokens for PIM Modula-2
+    pim_lexemes_to_reject = (
+        '!', '`', '@', '$', '%', '?', '\\', '==', '++', '--', '::', '*.',
+        '+>', '->', '<<', '>>', '|#', '##',
+    )
+
+    # PIM Modula-2 Additional Reserved Words Dataset
+    pim_additional_reserved_words = (
+        # 3 additional reserved words
+        'EXPORT', 'QUALIFIED', 'WITH',
+    )
+
+    # PIM Modula-2 Additional Builtins Dataset
+    pim_additional_builtins = (
+        # 16 additional builtins
+        'BITSET', 'CAP', 'DEC', 'DISPOSE', 'EXCL', 'FLOAT', 'HALT', 'HIGH',
+        'INC', 'INCL', 'NEW', 'NIL', 'PROC', 'SIZE', 'TRUNC', 'VAL',
+    )
+
+    # PIM Modula-2 Additional Pseudo-Module Builtins Dataset
+    pim_additional_pseudo_builtins = (
+        # 5 additional pseudo builtins
+        'SYSTEM', 'PROCESS', 'TSIZE', 'NEWPROCESS', 'TRANSFER',
+    )
+
+#  I S O   M o d u l a - 2   D a t a s e t s
+
+    # Lexemes to Mark as Error Tokens for ISO Modula-2
+    iso_lexemes_to_reject = (
+        '`', '$', '%', '?', '\\', '==', '++', '--', '::', '*.', '+>', '->',
+        '<<', '>>', '|#', '##',
+    )
+
+    # ISO Modula-2 Additional Reserved Words Dataset
+    iso_additional_reserved_words = (
+        # 9 additional reserved words (ISO 10514-1)
+        'EXCEPT', 'EXPORT', 'FINALLY', 'FORWARD', 'PACKEDSET', 'QUALIFIED',
+        'REM', 'RETRY', 'WITH',
+        # 10 additional reserved words (ISO 10514-2 & ISO 10514-3)
+        'ABSTRACT', 'AS', 'CLASS', 'GUARD', 'INHERIT', 'OVERRIDE', 'READONLY',
+        'REVEAL', 'TRACED', 'UNSAFEGUARDED',
+    )
+
+    # ISO Modula-2 Additional Builtins Dataset
+    iso_additional_builtins = (
+        # 26 additional builtins (ISO 10514-1)
+        'BITSET', 'CAP', 'CMPLX', 'COMPLEX', 'DEC', 'DISPOSE', 'EXCL', 'FLOAT',
+        'HALT', 'HIGH', 'IM', 'INC', 'INCL', 'INT', 'INTERRUPTIBLE',  'LENGTH',
+        'LFLOAT', 'LONGCOMPLEX', 'NEW', 'PROC', 'PROTECTION', 'RE', 'SIZE',
+        'TRUNC', 'UNINTERRUBTIBLE', 'VAL',
+        # 5 additional builtins (ISO 10514-2 & ISO 10514-3)
+        'CREATE', 'DESTROY', 'EMPTY', 'ISMEMBER', 'SELF',
+    )
+
+    # ISO Modula-2 Additional Pseudo-Module Builtins Dataset
+    iso_additional_pseudo_builtins = (
+        # 14 additional builtins (SYSTEM)
+        'SYSTEM', 'BITSPERLOC', 'LOCSPERBYTE', 'LOCSPERWORD', 'LOC',
+        'ADDADR', 'SUBADR', 'DIFADR', 'MAKEADR', 'ADR',
+        'ROTATE', 'SHIFT', 'CAST', 'TSIZE',
+        # 13 additional builtins (COROUTINES)
+        'COROUTINES', 'ATTACH', 'COROUTINE', 'CURRENT', 'DETACH', 'HANDLER',
+        'INTERRUPTSOURCE', 'IOTRANSFER', 'IsATTACHED', 'LISTEN',
+        'NEWCOROUTINE', 'PROT', 'TRANSFER',
+        # 9 additional builtins (EXCEPTIONS)
+        'EXCEPTIONS', 'AllocateSource', 'CurrentNumber', 'ExceptionNumber',
+        'ExceptionSource', 'GetMessage', 'IsCurrentSource',
+        'IsExceptionalExecution', 'RAISE',
+        # 3 additional builtins (TERMINATION)
+        'TERMINATION', 'IsTerminating', 'HasHalted',
+        # 4 additional builtins (M2EXCEPTION)
+        'M2EXCEPTION', 'M2Exceptions', 'M2Exception', 'IsM2Exception',
+        'indexException', 'rangeException', 'caseSelectException',
+        'invalidLocation', 'functionException', 'wholeValueException',
+        'wholeDivException', 'realValueException', 'realDivException',
+        'complexValueException', 'complexDivException', 'protException',
+        'sysException', 'coException', 'exException',
+    )
+
+#  M o d u l a - 2   R 1 0   D a t a s e t s
+
+    # Lexemes to Mark as Error Tokens for Modula-2 R10
+    m2r10_lexemes_to_reject = (
+        '!', '`', '@', '$', '%', '&', '<>',
+    )
+
+    # Modula-2 R10 reserved words in addition to the common set
+    m2r10_additional_reserved_words = (
+        # 12 additional reserved words
+        'ALIAS', 'ARGLIST', 'BLUEPRINT', 'COPY', 'GENLIB', 'INDETERMINATE',
+        'NEW', 'NONE', 'OPAQUE', 'REFERENTIAL', 'RELEASE', 'RETAIN',
+        # 2 additional reserved words with symbolic assembly option
+        'ASM', 'REG',
+    )
+
+    # Modula-2 R10 builtins in addition to the common set
+    m2r10_additional_builtins = (
+        # 26 additional builtins
+        'CARDINAL', 'COUNT', 'EMPTY', 'EXISTS', 'INSERT', 'LENGTH', 'LONGCARD',
+        'OCTET', 'PTR', 'PRED', 'READ', 'READNEW', 'REMOVE', 'RETRIEVE', 'SORT',
+        'STORE', 'SUBSET', 'SUCC', 'TLIMIT', 'TMAX', 'TMIN', 'TRUE', 'TSIZE',
+        'UNICHAR', 'WRITE', 'WRITEF',
+    )
+
+    # Modula-2 R10 Additional Pseudo-Module Builtins Dataset
+    m2r10_additional_pseudo_builtins = (
+        # 13 additional builtins (TPROPERTIES)
+        'TPROPERTIES', 'PROPERTY', 'LITERAL', 'TPROPERTY', 'TLITERAL',
+        'TBUILTIN', 'TDYN', 'TREFC', 'TNIL', 'TBASE', 'TPRECISION',
+        'TMAXEXP', 'TMINEXP',
+        # 4 additional builtins (CONVERSION)
+        'CONVERSION', 'TSXFSIZE', 'SXF', 'VAL',
+        # 35 additional builtins (UNSAFE)
+        'UNSAFE', 'CAST', 'INTRINSIC', 'AVAIL', 'ADD', 'SUB', 'ADDC', 'SUBC',
+        'FETCHADD', 'FETCHSUB', 'SHL', 'SHR', 'ASHR', 'ROTL', 'ROTR', 'ROTLC',
+        'ROTRC', 'BWNOT', 'BWAND', 'BWOR', 'BWXOR', 'BWNAND', 'BWNOR',
+        'SETBIT', 'TESTBIT', 'LSBIT', 'MSBIT', 'CSBITS', 'BAIL', 'HALT',
+        'TODO', 'FFI', 'ADDR', 'VARGLIST', 'VARGC',
+        # 11 additional builtins (ATOMIC)
+        'ATOMIC', 'INTRINSIC', 'AVAIL', 'SWAP', 'CAS', 'INC', 'DEC', 'BWAND',
+        'BWNAND', 'BWOR', 'BWXOR',
+        # 7 additional builtins (COMPILER)
+        'COMPILER', 'DEBUG', 'MODNAME', 'PROCNAME', 'LINENUM', 'DEFAULT',
+        'HASH',
+        # 5 additional builtins (ASSEMBLER)
+        'ASSEMBLER', 'REGISTER', 'SETREG', 'GETREG', 'CODE',
+    )
+
+#  O b j e c t i v e   M o d u l a - 2   D a t a s e t s
+
+    # Lexemes to Mark as Error Tokens for Objective Modula-2
+    objm2_lexemes_to_reject = (
+        '!', '$', '%', '&', '<>',
+    )
+
+    # Objective Modula-2 Extensions
+    # reserved words in addition to Modula-2 R10
+    objm2_additional_reserved_words = (
+        # 16 additional reserved words
+        'BYCOPY', 'BYREF', 'CLASS', 'CONTINUE', 'CRITICAL', 'INOUT', 'METHOD',
+        'ON', 'OPTIONAL', 'OUT', 'PRIVATE', 'PROTECTED', 'PROTOCOL', 'PUBLIC',
+        'SUPER', 'TRY',
+    )
+
+    # Objective Modula-2 Extensions
+    # builtins in addition to Modula-2 R10
+    objm2_additional_builtins = (
+        # 3 additional builtins
+        'OBJECT', 'NO', 'YES',
+    )
+
+    # Objective Modula-2 Extensions
+    # pseudo-module builtins in addition to Modula-2 R10
+    objm2_additional_pseudo_builtins = (
+        # None
+    )
+
+#  A g l e t   M o d u l a - 2   D a t a s e t s
+
+    # Aglet Extensions
+    # reserved words in addition to ISO Modula-2
+    aglet_additional_reserved_words = (
+        # None
+    )
+
+    # Aglet Extensions
+    # builtins in addition to ISO Modula-2
+    aglet_additional_builtins = (
+        # 9 additional builtins
+        'BITSET8', 'BITSET16', 'BITSET32', 'CARDINAL8', 'CARDINAL16',
+        'CARDINAL32', 'INTEGER8', 'INTEGER16', 'INTEGER32',
+    )
+
+    # Aglet Modula-2 Extensions
+    # pseudo-module builtins in addition to ISO Modula-2
+    aglet_additional_pseudo_builtins = (
+        # None
+    )
+
+#  G N U   M o d u l a - 2   D a t a s e t s
+
+    # GNU Extensions
+    # reserved words in addition to PIM Modula-2
+    gm2_additional_reserved_words = (
+        # 10 additional reserved words
+        'ASM', '__ATTRIBUTE__', '__BUILTIN__', '__COLUMN__', '__DATE__',
+        '__FILE__', '__FUNCTION__', '__LINE__', '__MODULE__', 'VOLATILE',
+    )
+
+    # GNU Extensions
+    # builtins in addition to PIM Modula-2
+    gm2_additional_builtins = (
+        # 21 additional builtins
+        'BITSET8', 'BITSET16', 'BITSET32', 'CARDINAL8', 'CARDINAL16',
+        'CARDINAL32', 'CARDINAL64', 'COMPLEX32', 'COMPLEX64', 'COMPLEX96',
+        'COMPLEX128', 'INTEGER8', 'INTEGER16', 'INTEGER32', 'INTEGER64',
+        'REAL8', 'REAL16', 'REAL32', 'REAL96', 'REAL128', 'THROW',
+    )
+
+    # GNU Extensions
+    # pseudo-module builtins in addition to PIM Modula-2
+    gm2_additional_pseudo_builtins = (
+        # None
+    )
+
+#  p 1   M o d u l a - 2   D a t a s e t s
+
+    # p1 Extensions
+    # reserved words in addition to ISO Modula-2
+    p1_additional_reserved_words = (
+        # None
+    )
+
+    # p1 Extensions
+    # builtins in addition to ISO Modula-2
+    p1_additional_builtins = (
+        # None
+    )
+
+    # p1 Modula-2 Extensions
+    # pseudo-module builtins in addition to ISO Modula-2
+    p1_additional_pseudo_builtins = (
+        # 1 additional builtin
+        'BCD',
+    )
+
+#  X D S   M o d u l a - 2   D a t a s e t s
+
+    # XDS Extensions
+    # reserved words in addition to ISO Modula-2
+    xds_additional_reserved_words = (
+        # 1 additional reserved word
+        'SEQ',
+    )
+
+    # XDS Extensions
+    # builtins in addition to ISO Modula-2
+    xds_additional_builtins = (
+        # 9 additional builtins
+        'ASH', 'ASSERT', 'DIFFADR_TYPE', 'ENTIER', 'INDEX', 'LEN',
+        'LONGCARD', 'SHORTCARD', 'SHORTINT',
+    )
+
+    # XDS Modula-2 Extensions
+    # pseudo-module builtins in addition to ISO Modula-2
+    xds_additional_pseudo_builtins = (
+        # 22 additional builtins (SYSTEM)
+        'PROCESS', 'NEWPROCESS', 'BOOL8', 'BOOL16', 'BOOL32', 'CARD8',
+        'CARD16', 'CARD32', 'INT8', 'INT16', 'INT32', 'REF', 'MOVE',
+        'FILL', 'GET', 'PUT', 'CC', 'int', 'unsigned', 'size_t', 'void'
+        # 3 additional builtins (COMPILER)
+        'COMPILER', 'OPTION', 'EQUATION'
+    )
+
+#  P I M   S t a n d a r d   L i b r a r y   D a t a s e t s
+
+    # PIM Modula-2 Standard Library Modules Dataset
+    pim_stdlib_module_identifiers = (
+        'Terminal', 'FileSystem', 'InOut', 'RealInOut', 'MathLib0', 'Storage',
+    )
+
+    # PIM Modula-2 Standard Library Types Dataset
+    pim_stdlib_type_identifiers = (
+        'Flag', 'FlagSet', 'Response', 'Command', 'Lock', 'Permission',
+        'MediumType', 'File', 'FileProc', 'DirectoryProc', 'FileCommand',
+        'DirectoryCommand',
+    )
+
+    # PIM Modula-2 Standard Library Procedures Dataset
+    pim_stdlib_proc_identifiers = (
+        'Read', 'BusyRead', 'ReadAgain', 'Write', 'WriteString', 'WriteLn',
+        'Create', 'Lookup', 'Close', 'Delete', 'Rename', 'SetRead', 'SetWrite',
+        'SetModify', 'SetOpen', 'Doio', 'SetPos', 'GetPos', 'Length', 'Reset',
+        'Again', 'ReadWord', 'WriteWord', 'ReadChar', 'WriteChar',
+        'CreateMedium', 'DeleteMedium', 'AssignName', 'DeassignName',
+        'ReadMedium', 'LookupMedium', 'OpenInput', 'OpenOutput', 'CloseInput',
+        'CloseOutput', 'ReadString', 'ReadInt', 'ReadCard', 'ReadWrd',
+        'WriteInt', 'WriteCard', 'WriteOct', 'WriteHex', 'WriteWrd',
+        'ReadReal', 'WriteReal', 'WriteFixPt', 'WriteRealOct', 'sqrt', 'exp',
+        'ln', 'sin', 'cos', 'arctan', 'entier', 'ALLOCATE', 'DEALLOCATE',
+    )
+
+    # PIM Modula-2 Standard Library Variables Dataset
+    pim_stdlib_var_identifiers = (
+        'Done', 'termCH', 'in', 'out'
+    )
+
+    # PIM Modula-2 Standard Library Constants Dataset
+    pim_stdlib_const_identifiers = (
+        'EOL',
+    )
+
+#  I S O   S t a n d a r d   L i b r a r y   D a t a s e t s
+
+    # ISO Modula-2 Standard Library Modules Dataset
+    iso_stdlib_module_identifiers = (
+        # TO DO
+    )
+
+    # ISO Modula-2 Standard Library Types Dataset
+    iso_stdlib_type_identifiers = (
+        # TO DO
+    )
+
+    # ISO Modula-2 Standard Library Procedures Dataset
+    iso_stdlib_proc_identifiers = (
+        # TO DO
+    )
+
+    # ISO Modula-2 Standard Library Variables Dataset
+    iso_stdlib_var_identifiers = (
+        # TO DO
+    )
+
+    # ISO Modula-2 Standard Library Constants Dataset
+    iso_stdlib_const_identifiers = (
+        # TO DO
+    )
+
+#  M 2   R 1 0   S t a n d a r d   L i b r a r y   D a t a s e t s
+
+    # Modula-2 R10 Standard Library ADTs Dataset
+    m2r10_stdlib_adt_identifiers = (
+        'BCD', 'LONGBCD', 'BITSET', 'SHORTBITSET', 'LONGBITSET',
+        'LONGLONGBITSET', 'COMPLEX', 'LONGCOMPLEX', 'SHORTCARD', 'LONGLONGCARD',
+        'SHORTINT', 'LONGLONGINT', 'POSINT', 'SHORTPOSINT', 'LONGPOSINT',
+        'LONGLONGPOSINT', 'BITSET8', 'BITSET16', 'BITSET32', 'BITSET64',
+        'BITSET128', 'BS8', 'BS16', 'BS32', 'BS64', 'BS128', 'CARDINAL8',
+        'CARDINAL16', 'CARDINAL32', 'CARDINAL64', 'CARDINAL128', 'CARD8',
+        'CARD16', 'CARD32', 'CARD64', 'CARD128', 'INTEGER8', 'INTEGER16',
+        'INTEGER32', 'INTEGER64', 'INTEGER128', 'INT8', 'INT16', 'INT32',
+        'INT64', 'INT128', 'STRING', 'UNISTRING',
+    )
+
+    # Modula-2 R10 Standard Library Blueprints Dataset
+    m2r10_stdlib_blueprint_identifiers = (
+        'ProtoRoot', 'ProtoComputational', 'ProtoNumeric', 'ProtoScalar',
+        'ProtoNonScalar', 'ProtoCardinal', 'ProtoInteger', 'ProtoReal',
+        'ProtoComplex', 'ProtoVector', 'ProtoTuple', 'ProtoCompArray',
+        'ProtoCollection', 'ProtoStaticArray', 'ProtoStaticSet',
+        'ProtoStaticString', 'ProtoArray', 'ProtoString', 'ProtoSet',
+        'ProtoMultiSet', 'ProtoDictionary', 'ProtoMultiDict', 'ProtoExtension',
+        'ProtoIO', 'ProtoCardMath', 'ProtoIntMath', 'ProtoRealMath',
+    )
+
+    # Modula-2 R10 Standard Library Modules Dataset
+    m2r10_stdlib_module_identifiers = (
+        'ASCII', 'BooleanIO', 'CharIO', 'UnicharIO', 'OctetIO',
+        'CardinalIO', 'LongCardIO', 'IntegerIO', 'LongIntIO', 'RealIO',
+        'LongRealIO', 'BCDIO', 'LongBCDIO', 'CardMath', 'LongCardMath',
+        'IntMath', 'LongIntMath', 'RealMath', 'LongRealMath', 'BCDMath',
+        'LongBCDMath', 'FileIO', 'FileSystem', 'Storage', 'IOSupport',
+    )
+
+    # Modula-2 R10 Standard Library Types Dataset
+    m2r10_stdlib_type_identifiers = (
+        'File', 'Status',
+        # TO BE COMPLETED
+    )
+
+    # Modula-2 R10 Standard Library Procedures Dataset
+    m2r10_stdlib_proc_identifiers = (
+        'ALLOCATE', 'DEALLOCATE', 'SIZE',
+        # TO BE COMPLETED
+    )
+
+    # Modula-2 R10 Standard Library Variables Dataset
+    m2r10_stdlib_var_identifiers = (
+        'stdIn', 'stdOut', 'stdErr',
+    )
+
+    # Modula-2 R10 Standard Library Constants Dataset
+    m2r10_stdlib_const_identifiers = (
+        'pi', 'tau',
+    )
+
+#  D i a l e c t s
+
+    # Dialect modes
+    dialects = (
+        'unknown',
+        'm2pim', 'm2iso', 'm2r10', 'objm2',
+        'm2iso+aglet', 'm2pim+gm2', 'm2iso+p1', 'm2iso+xds',
+    )
+
+#   D a t a b a s e s
+
+    # Lexemes to Mark as Errors Database
+    lexemes_to_reject_db = {
+        # Lexemes to reject for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Lexemes to reject for PIM Modula-2
+        'm2pim': (
+            pim_lexemes_to_reject,
+        ),
+        # Lexemes to reject for ISO Modula-2
+        'm2iso': (
+            iso_lexemes_to_reject,
+        ),
+        # Lexemes to reject for Modula-2 R10
+        'm2r10': (
+            m2r10_lexemes_to_reject,
+        ),
+        # Lexemes to reject for Objective Modula-2
+        'objm2': (
+            objm2_lexemes_to_reject,
+        ),
+        # Lexemes to reject for Aglet Modula-2
+        'm2iso+aglet': (
+            iso_lexemes_to_reject,
+        ),
+        # Lexemes to reject for GNU Modula-2
+        'm2pim+gm2': (
+            pim_lexemes_to_reject,
+        ),
+        # Lexemes to reject for p1 Modula-2
+        'm2iso+p1': (
+            iso_lexemes_to_reject,
+        ),
+        # Lexemes to reject for XDS Modula-2
+        'm2iso+xds': (
+            iso_lexemes_to_reject,
+        ),
+    }
+
+    # Reserved Words Database
+    reserved_words_db = {
+        # Reserved words for unknown dialect
+        'unknown': (
+            common_reserved_words,
+            pim_additional_reserved_words,
+            iso_additional_reserved_words,
+            m2r10_additional_reserved_words,
+        ),
+
+        # Reserved words for PIM Modula-2
+        'm2pim': (
+            common_reserved_words,
+            pim_additional_reserved_words,
+        ),
+
+        # Reserved words for Modula-2 R10
+        'm2iso': (
+            common_reserved_words,
+            iso_additional_reserved_words,
+        ),
+
+        # Reserved words for ISO Modula-2
+        'm2r10': (
+            common_reserved_words,
+            m2r10_additional_reserved_words,
+        ),
+
+        # Reserved words for Objective Modula-2
+        'objm2': (
+            common_reserved_words,
+            m2r10_additional_reserved_words,
+            objm2_additional_reserved_words,
+        ),
+
+        # Reserved words for Aglet Modula-2 Extensions
+        'm2iso+aglet': (
+            common_reserved_words,
+            iso_additional_reserved_words,
+            aglet_additional_reserved_words,
+        ),
+
+        # Reserved words for GNU Modula-2 Extensions
+        'm2pim+gm2': (
+            common_reserved_words,
+            pim_additional_reserved_words,
+            gm2_additional_reserved_words,
+        ),
+
+        # Reserved words for p1 Modula-2 Extensions
+        'm2iso+p1': (
+            common_reserved_words,
+            iso_additional_reserved_words,
+            p1_additional_reserved_words,
+        ),
+
+        # Reserved words for XDS Modula-2 Extensions
+        'm2iso+xds': (
+            common_reserved_words,
+            iso_additional_reserved_words,
+            xds_additional_reserved_words,
+        ),
+    }
+
+    # Builtins Database
+    builtins_db = {
+        # Builtins for unknown dialect
+        'unknown': (
+            common_builtins,
+            pim_additional_builtins,
+            iso_additional_builtins,
+            m2r10_additional_builtins,
+        ),
+
+        # Builtins for PIM Modula-2
+        'm2pim': (
+            common_builtins,
+            pim_additional_builtins,
+        ),
+
+        # Builtins for ISO Modula-2
+        'm2iso': (
+            common_builtins,
+            iso_additional_builtins,
+        ),
+
+        # Builtins for ISO Modula-2
+        'm2r10': (
+            common_builtins,
+            m2r10_additional_builtins,
+        ),
+
+        # Builtins for Objective Modula-2
+        'objm2': (
+            common_builtins,
+            m2r10_additional_builtins,
+            objm2_additional_builtins,
+        ),
+
+        # Builtins for Aglet Modula-2 Extensions
+        'm2iso+aglet': (
+            common_builtins,
+            iso_additional_builtins,
+            aglet_additional_builtins,
+        ),
+
+        # Builtins for GNU Modula-2 Extensions
+        'm2pim+gm2': (
+            common_builtins,
+            pim_additional_builtins,
+            gm2_additional_builtins,
+        ),
+
+        # Builtins for p1 Modula-2 Extensions
+        'm2iso+p1': (
+            common_builtins,
+            iso_additional_builtins,
+            p1_additional_builtins,
+        ),
+
+        # Builtins for XDS Modula-2 Extensions
+        'm2iso+xds': (
+            common_builtins,
+            iso_additional_builtins,
+            xds_additional_builtins,
+        ),
+    }
+
+    # Pseudo-Module Builtins Database
+    pseudo_builtins_db = {
+        # Builtins for unknown dialect
+        'unknown': (
+            common_pseudo_builtins,
+            pim_additional_pseudo_builtins,
+            iso_additional_pseudo_builtins,
+            m2r10_additional_pseudo_builtins,
+        ),
+
+        # Builtins for PIM Modula-2
+        'm2pim': (
+            common_pseudo_builtins,
+            pim_additional_pseudo_builtins,
+        ),
+
+        # Builtins for ISO Modula-2
+        'm2iso': (
+            common_pseudo_builtins,
+            iso_additional_pseudo_builtins,
+        ),
+
+        # Builtins for ISO Modula-2
+        'm2r10': (
+            common_pseudo_builtins,
+            m2r10_additional_pseudo_builtins,
+        ),
+
+        # Builtins for Objective Modula-2
+        'objm2': (
+            common_pseudo_builtins,
+            m2r10_additional_pseudo_builtins,
+            objm2_additional_pseudo_builtins,
+        ),
+
+        # Builtins for Aglet Modula-2 Extensions
+        'm2iso+aglet': (
+            common_pseudo_builtins,
+            iso_additional_pseudo_builtins,
+            aglet_additional_pseudo_builtins,
+        ),
+
+        # Builtins for GNU Modula-2 Extensions
+        'm2pim+gm2': (
+            common_pseudo_builtins,
+            pim_additional_pseudo_builtins,
+            gm2_additional_pseudo_builtins,
+        ),
+
+        # Builtins for p1 Modula-2 Extensions
+        'm2iso+p1': (
+            common_pseudo_builtins,
+            iso_additional_pseudo_builtins,
+            p1_additional_pseudo_builtins,
+        ),
+
+        # Builtins for XDS Modula-2 Extensions
+        'm2iso+xds': (
+            common_pseudo_builtins,
+            iso_additional_pseudo_builtins,
+            xds_additional_pseudo_builtins,
+        ),
+    }
+
+    # Standard Library ADTs Database
+    stdlib_adts_db = {
+        # Empty entry for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Standard Library ADTs for PIM Modula-2
+        'm2pim': (
+            # No first class library types
+        ),
+
+        # Standard Library ADTs for ISO Modula-2
+        'm2iso': (
+            # No first class library types
+        ),
+
+        # Standard Library ADTs for Modula-2 R10
+        'm2r10': (
+            m2r10_stdlib_adt_identifiers,
+        ),
+
+        # Standard Library ADTs for Objective Modula-2
+        'objm2': (
+            m2r10_stdlib_adt_identifiers,
+        ),
+
+        # Standard Library ADTs for Aglet Modula-2
+        'm2iso+aglet': (
+            # No first class library types
+        ),
+
+        # Standard Library ADTs for GNU Modula-2
+        'm2pim+gm2': (
+            # No first class library types
+        ),
+
+        # Standard Library ADTs for p1 Modula-2
+        'm2iso+p1': (
+            # No first class library types
+        ),
+
+        # Standard Library ADTs for XDS Modula-2
+        'm2iso+xds': (
+            # No first class library types
+        ),
+    }
+
+    # Standard Library Modules Database
+    stdlib_modules_db = {
+        # Empty entry for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Standard Library Modules for PIM Modula-2
+        'm2pim': (
+            pim_stdlib_module_identifiers,
+        ),
+
+        # Standard Library Modules for ISO Modula-2
+        'm2iso': (
+            iso_stdlib_module_identifiers,
+        ),
+
+        # Standard Library Modules for Modula-2 R10
+        'm2r10': (
+            m2r10_stdlib_blueprint_identifiers,
+            m2r10_stdlib_module_identifiers,
+            m2r10_stdlib_adt_identifiers,
+        ),
+
+        # Standard Library Modules for Objective Modula-2
+        'objm2': (
+            m2r10_stdlib_blueprint_identifiers,
+            m2r10_stdlib_module_identifiers,
+        ),
+
+        # Standard Library Modules for Aglet Modula-2
+        'm2iso+aglet': (
+            iso_stdlib_module_identifiers,
+        ),
+
+        # Standard Library Modules for GNU Modula-2
+        'm2pim+gm2': (
+            pim_stdlib_module_identifiers,
+        ),
+
+        # Standard Library Modules for p1 Modula-2
+        'm2iso+p1': (
+            iso_stdlib_module_identifiers,
+        ),
+
+        # Standard Library Modules for XDS Modula-2
+        'm2iso+xds': (
+            iso_stdlib_module_identifiers,
+        ),
+    }
+
+    # Standard Library Types Database
+    stdlib_types_db = {
+        # Empty entry for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Standard Library Types for PIM Modula-2
+        'm2pim': (
+            pim_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for ISO Modula-2
+        'm2iso': (
+            iso_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for Modula-2 R10
+        'm2r10': (
+            m2r10_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for Objective Modula-2
+        'objm2': (
+            m2r10_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for Aglet Modula-2
+        'm2iso+aglet': (
+            iso_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for GNU Modula-2
+        'm2pim+gm2': (
+            pim_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for p1 Modula-2
+        'm2iso+p1': (
+            iso_stdlib_type_identifiers,
+        ),
+
+        # Standard Library Types for XDS Modula-2
+        'm2iso+xds': (
+            iso_stdlib_type_identifiers,
+        ),
+    }
+
+    # Standard Library Procedures Database
+    stdlib_procedures_db = {
+        # Empty entry for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Standard Library Procedures for PIM Modula-2
+        'm2pim': (
+            pim_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for ISO Modula-2
+        'm2iso': (
+            iso_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for Modula-2 R10
+        'm2r10': (
+            m2r10_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for Objective Modula-2
+        'objm2': (
+            m2r10_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for Aglet Modula-2
+        'm2iso+aglet': (
+            iso_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for GNU Modula-2
+        'm2pim+gm2': (
+            pim_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for p1 Modula-2
+        'm2iso+p1': (
+            iso_stdlib_proc_identifiers,
+        ),
+
+        # Standard Library Procedures for XDS Modula-2
+        'm2iso+xds': (
+            iso_stdlib_proc_identifiers,
+        ),
+    }
+
+    # Standard Library Variables Database
+    stdlib_variables_db = {
+        # Empty entry for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Standard Library Variables for PIM Modula-2
+        'm2pim': (
+            pim_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for ISO Modula-2
+        'm2iso': (
+            iso_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for Modula-2 R10
+        'm2r10': (
+            m2r10_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for Objective Modula-2
+        'objm2': (
+            m2r10_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for Aglet Modula-2
+        'm2iso+aglet': (
+            iso_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for GNU Modula-2
+        'm2pim+gm2': (
+            pim_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for p1 Modula-2
+        'm2iso+p1': (
+            iso_stdlib_var_identifiers,
+        ),
+
+        # Standard Library Variables for XDS Modula-2
+        'm2iso+xds': (
+            iso_stdlib_var_identifiers,
+        ),
+    }
+
+    # Standard Library Constants Database
+    stdlib_constants_db = {
+        # Empty entry for unknown dialect
+        'unknown': (
+            # LEAVE THIS EMPTY
+        ),
+        # Standard Library Constants for PIM Modula-2
+        'm2pim': (
+            pim_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for ISO Modula-2
+        'm2iso': (
+            iso_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for Modula-2 R10
+        'm2r10': (
+            m2r10_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for Objective Modula-2
+        'objm2': (
+            m2r10_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for Aglet Modula-2
+        'm2iso+aglet': (
+            iso_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for GNU Modula-2
+        'm2pim+gm2': (
+            pim_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for p1 Modula-2
+        'm2iso+p1': (
+            iso_stdlib_const_identifiers,
+        ),
+
+        # Standard Library Constants for XDS Modula-2
+        'm2iso+xds': (
+            iso_stdlib_const_identifiers,
+        ),
+    }
+
+#   M e t h o d s
+
+    # initialise a lexer instance
+    def __init__(self, **options):
+        #
+        # check dialect options
+        #
+        dialects = get_list_opt(options, 'dialect', [])
+        #
+        for dialect_option in dialects:
+            if dialect_option in self.dialects[1:-1]:
+                # valid dialect option found
+                self.set_dialect(dialect_option)
+                break
+        #
+        # Fallback Mode (DEFAULT)
+        else:
+            # no valid dialect option
+            self.set_dialect('unknown')
+        #
+        self.dialect_set_by_tag = False
+        #
+        # check style options
+        #
+        styles = get_list_opt(options, 'style', [])
+        #
+        # use lowercase mode for Algol style
+        if 'algol' in styles or 'algol_nu' in styles:
+            self.algol_publication_mode = True
+        else:
+            self.algol_publication_mode = False
+        #
+        # Check option flags
+        #
+        self.treat_stdlib_adts_as_builtins = get_bool_opt(
+            options, 'treat_stdlib_adts_as_builtins', True)
+        #
+        # call superclass initialiser
+        RegexLexer.__init__(self, **options)
+
+    # Set lexer to a specified dialect
+    def set_dialect(self, dialect_id):
+        #
+        # if __debug__:
+        #    print 'entered set_dialect with arg: ', dialect_id
+        #
+        # check dialect name against known dialects
+        if dialect_id not in self.dialects:
+            dialect = 'unknown'  # default
+        else:
+            dialect = dialect_id
+        #
+        # compose lexemes to reject set
+        lexemes_to_reject_set = set()
+        # add each list of reject lexemes for this dialect
+        for list in self.lexemes_to_reject_db[dialect]:
+            lexemes_to_reject_set.update(set(list))
+        #
+        # compose reserved words set
+        reswords_set = set()
+        # add each list of reserved words for this dialect
+        for list in self.reserved_words_db[dialect]:
+            reswords_set.update(set(list))
+        #
+        # compose builtins set
+        builtins_set = set()
+        # add each list of builtins for this dialect excluding reserved words
+        for list in self.builtins_db[dialect]:
+            builtins_set.update(set(list).difference(reswords_set))
+        #
+        # compose pseudo-builtins set
+        pseudo_builtins_set = set()
+        # add each list of builtins for this dialect excluding reserved words
+        for list in self.pseudo_builtins_db[dialect]:
+            pseudo_builtins_set.update(set(list).difference(reswords_set))
+        #
+        # compose ADTs set
+        adts_set = set()
+        # add each list of ADTs for this dialect excluding reserved words
+        for list in self.stdlib_adts_db[dialect]:
+            adts_set.update(set(list).difference(reswords_set))
+        #
+        # compose modules set
+        modules_set = set()
+        # add each list of builtins for this dialect excluding builtins
+        for list in self.stdlib_modules_db[dialect]:
+            modules_set.update(set(list).difference(builtins_set))
+        #
+        # compose types set
+        types_set = set()
+        # add each list of types for this dialect excluding builtins
+        for list in self.stdlib_types_db[dialect]:
+            types_set.update(set(list).difference(builtins_set))
+        #
+        # compose procedures set
+        procedures_set = set()
+        # add each list of procedures for this dialect excluding builtins
+        for list in self.stdlib_procedures_db[dialect]:
+            procedures_set.update(set(list).difference(builtins_set))
+        #
+        # compose variables set
+        variables_set = set()
+        # add each list of variables for this dialect excluding builtins
+        for list in self.stdlib_variables_db[dialect]:
+            variables_set.update(set(list).difference(builtins_set))
+        #
+        # compose constants set
+        constants_set = set()
+        # add each list of constants for this dialect excluding builtins
+        for list in self.stdlib_constants_db[dialect]:
+            constants_set.update(set(list).difference(builtins_set))
+        #
+        # update lexer state
+        self.dialect = dialect
+        self.lexemes_to_reject = lexemes_to_reject_set
+        self.reserved_words = reswords_set
+        self.builtins = builtins_set
+        self.pseudo_builtins = pseudo_builtins_set
+        self.adts = adts_set
+        self.modules = modules_set
+        self.types = types_set
+        self.procedures = procedures_set
+        self.variables = variables_set
+        self.constants = constants_set
+        #
+        # if __debug__:
+        #    print 'exiting set_dialect'
+        #    print ' self.dialect: ', self.dialect
+        #    print ' self.lexemes_to_reject: ', self.lexemes_to_reject
+        #    print ' self.reserved_words: ', self.reserved_words
+        #    print ' self.builtins: ', self.builtins
+        #    print ' self.pseudo_builtins: ', self.pseudo_builtins
+        #    print ' self.adts: ', self.adts
+        #    print ' self.modules: ', self.modules
+        #    print ' self.types: ', self.types
+        #    print ' self.procedures: ', self.procedures
+        #    print ' self.variables: ', self.variables
+        #    print ' self.types: ', self.types
+        #    print ' self.constants: ', self.constants
+
+    # Extracts a dialect name from a dialect tag comment string  and checks
+    # the extracted name against known dialects.  If a match is found,  the
+    # matching name is returned, otherwise dialect id 'unknown' is returned
+    def get_dialect_from_dialect_tag(self, dialect_tag):
+        #
+        # if __debug__:
+        #    print 'entered get_dialect_from_dialect_tag with arg: ', dialect_tag
+        #
+        # constants
+        left_tag_delim = '(*!'
+        right_tag_delim = '*)'
+        left_tag_delim_len = len(left_tag_delim)
+        right_tag_delim_len = len(right_tag_delim)
+        indicator_start = left_tag_delim_len
+        indicator_end = -(right_tag_delim_len)
+        #
+        # check comment string for dialect indicator
+        if len(dialect_tag) > (left_tag_delim_len + right_tag_delim_len) \
+           and dialect_tag.startswith(left_tag_delim) \
+           and dialect_tag.endswith(right_tag_delim):
+            #
+            # if __debug__:
+            #    print 'dialect tag found'
+            #
+            # extract dialect indicator
+            indicator = dialect_tag[indicator_start:indicator_end]
+            #
+            # if __debug__:
+            #    print 'extracted: ', indicator
+            #
+            # check against known dialects
+            for index in range(1, len(self.dialects)):
+                #
+                # if __debug__:
+                #    print 'dialects[', index, ']: ', self.dialects[index]
+                #
+                if indicator == self.dialects[index]:
+                    #
+                    # if __debug__:
+                    #    print 'matching dialect found'
+                    #
+                    # indicator matches known dialect
+                    return indicator
+            else:
+                # indicator does not match any dialect
+                return 'unknown'  # default
+        else:
+            # invalid indicator string
+            return 'unknown'  # default
+
+    # intercept the token stream, modify token attributes and return them
+    def get_tokens_unprocessed(self, text):
+        for index, token, value in RegexLexer.get_tokens_unprocessed(self, text):
+            #
+            # check for dialect tag if dialect has not been set by tag
+            if not self.dialect_set_by_tag and token == Comment.Special:
+                indicated_dialect = self.get_dialect_from_dialect_tag(value)
+                if indicated_dialect != 'unknown':
+                    # token is a dialect indicator
+                    # reset reserved words and builtins
+                    self.set_dialect(indicated_dialect)
+                    self.dialect_set_by_tag = True
+            #
+            # check for reserved words, predefined and stdlib identifiers
+            if token is Name:
+                if value in self.reserved_words:
+                    token = Keyword.Reserved
+                    if self.algol_publication_mode:
+                        value = value.lower()
+                #
+                elif value in self.builtins:
+                    token = Name.Builtin
+                    if self.algol_publication_mode:
+                        value = value.lower()
+                #
+                elif value in self.pseudo_builtins:
+                    token = Name.Builtin.Pseudo
+                    if self.algol_publication_mode:
+                        value = value.lower()
+                #
+                elif value in self.adts:
+                    if not self.treat_stdlib_adts_as_builtins:
+                        token = Name.Namespace
+                    else:
+                        token = Name.Builtin.Pseudo
+                        if self.algol_publication_mode:
+                            value = value.lower()
+                #
+                elif value in self.modules:
+                    token = Name.Namespace
+                #
+                elif value in self.types:
+                    token = Name.Class
+                #
+                elif value in self.procedures:
+                    token = Name.Function
+                #
+                elif value in self.variables:
+                    token = Name.Variable
+                #
+                elif value in self.constants:
+                    token = Name.Constant
+            #
+            elif token in Number:
+                #
+                # mark prefix number literals as error for PIM and ISO dialects
+                if self.dialect not in ('unknown', 'm2r10', 'objm2'):
+                    if "'" in value or value[0:2] in ('0b', '0x', '0u'):
+                        token = Error
+                #
+                elif self.dialect in ('m2r10', 'objm2'):
+                    # mark base-8 number literals as errors for M2 R10 and ObjM2
+                    if token is Number.Oct:
+                        token = Error
+                    # mark suffix base-16 literals as errors for M2 R10 and ObjM2
+                    elif token is Number.Hex and 'H' in value:
+                        token = Error
+                    # mark real numbers with E as errors for M2 R10 and ObjM2
+                    elif token is Number.Float and 'E' in value:
+                        token = Error
+            #
+            elif token in Comment:
+                #
+                # mark single line comment as error for PIM and ISO dialects
+                if token is Comment.Single:
+                    if self.dialect not in ('unknown', 'm2r10', 'objm2'):
+                        token = Error
+                #
+                if token is Comment.Preproc:
+                    # mark ISO pragma as error for PIM dialects
+                    if value.startswith('<*') and \
+                       self.dialect.startswith('m2pim'):
+                        token = Error
+                    # mark PIM pragma as comment for other dialects
+                    elif value.startswith('(*$') and \
+                            self.dialect != 'unknown' and \
+                            not self.dialect.startswith('m2pim'):
+                        token = Comment.Multiline
+            #
+            else:  # token is neither Name nor Comment
+                #
+                # mark lexemes matching the dialect's error token set as errors
+                if value in self.lexemes_to_reject:
+                    token = Error
+                #
+                # substitute lexemes when in Algol mode
+                if self.algol_publication_mode:
+                    if value == '#':
+                        value = '≠'
+                    elif value == '<=':
+                        value = '≤'
+                    elif value == '>=':
+                        value = '≥'
+                    elif value == '==':
+                        value = '≡'
+                    elif value == '*.':
+                        value = '•'
+
+            # return result
+            yield index, token, value
+
+    def analyse_text(text):
+        """It's Pascal-like, but does not use FUNCTION -- uses PROCEDURE
+        instead."""
+
+        # Check if this looks like Pascal, if not, bail out early
+        if not ('(*' in text and '*)' in text and ':=' in text):
+            return
+
+        result = 0
+        # Procedure is in Modula2
+        if re.search(r'\bPROCEDURE\b', text):
+            result += 0.6
+
+        # FUNCTION is only valid in Pascal, but not in Modula2
+        if re.search(r'\bFUNCTION\b', text):
+            result = 0.0
+
+        return result

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/pygments/lexers/smithy.py

@@ -0,0 +1,77 @@
+"""
+    pygments.lexers.smithy
+    ~~~~~~~~~~~~~~~~~~~~~~
+
+    Lexers for the Smithy IDL.
+
+    :copyright: Copyright 2006-present by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+from pygments.lexer import RegexLexer, bygroups, words
+from pygments.token import Text, Comment, Keyword, Name, String, \
+    Number, Whitespace, Punctuation
+
+__all__ = ['SmithyLexer']
+
+
+class SmithyLexer(RegexLexer):
+    """
+    For Smithy IDL
+    """
+    name = 'Smithy'
+    url = 'https://awslabs.github.io/smithy/'
+    filenames = ['*.smithy']
+    aliases = ['smithy']
+    version_added = '2.10'
+
+    unquoted = r'[A-Za-z0-9_\.#$-]+'
+    identifier = r"[A-Za-z0-9_\.#$-]+"
+
+    simple_shapes = (
+        'use', 'byte', 'short', 'integer', 'long', 'float', 'document',
+        'double', 'bigInteger', 'bigDecimal', 'boolean', 'blob', 'string',
+        'timestamp',
+    )
+
+    aggregate_shapes = (
+       'apply', 'list', 'map', 'set', 'structure', 'union', 'resource',
+       'operation', 'service', 'trait'
+    )
+
+    tokens = {
+        'root': [
+            (r'///.*$', Comment.Multiline),
+            (r'//.*$', Comment),
+            (r'@[0-9a-zA-Z\.#-]*', Name.Decorator),
+            (r'(=)', Name.Decorator),
+            (r'^(\$version)(:)(.+)',
+                bygroups(Keyword.Declaration, Name.Decorator, Name.Class)),
+            (r'^(namespace)(\s+' + identifier + r')\b',
+                bygroups(Keyword.Declaration, Name.Class)),
+            (words(simple_shapes,
+                   prefix=r'^', suffix=r'(\s+' + identifier + r')\b'),
+                bygroups(Keyword.Declaration, Name.Class)),
+            (words(aggregate_shapes,
+                   prefix=r'^', suffix=r'(\s+' + identifier + r')'),
+                bygroups(Keyword.Declaration, Name.Class)),
+            (r'^(metadata)(\s+)((?:\S+)|(?:\"[^"]+\"))(\s*)(=)',
+                bygroups(Keyword.Declaration, Whitespace, Name.Class,
+                         Whitespace, Name.Decorator)),
+            (r"(true|false|null)", Keyword.Constant),
+            (r"(-?(?:0|[1-9]\d*)(?:\.\d+)?(?:[eE][+-]?\d+)?)", Number),
+            (identifier + ":", Name.Label),
+            (identifier, Name.Variable.Class),
+            (r'\[', Text, "#push"),
+            (r'\]', Text, "#pop"),
+            (r'\(', Text, "#push"),
+            (r'\)', Text, "#pop"),
+            (r'\{', Text, "#push"),
+            (r'\}', Text, "#pop"),
+            (r'"{3}(\\\\|\n|\\")*"{3}', String.Doc),
+            (r'"(\\\\|\n|\\"|[^"])*"', String.Double),
+            (r"'(\\\\|\n|\\'|[^'])*'", String.Single),
+            (r'[:,]+', Punctuation),
+            (r'\s+', Whitespace),
+        ]
+    }

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/pygments/lexers/textedit.py

@@ -0,0 +1,205 @@
+"""
+    pygments.lexers.textedit
+    ~~~~~~~~~~~~~~~~~~~~~~~~
+
+    Lexers for languages related to text processing.
+
+    :copyright: Copyright 2006-present by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+import re
+from bisect import bisect
+
+from pygments.lexer import RegexLexer, bygroups, default, include, this, using
+from pygments.lexers.python import PythonLexer
+from pygments.token import Comment, Keyword, Name, Number, Operator, \
+    Punctuation, String, Text, Whitespace
+
+__all__ = ['AwkLexer', 'SedLexer', 'VimLexer']
+
+
+class AwkLexer(RegexLexer):
+    """
+    For Awk scripts.
+    """
+
+    name = 'Awk'
+    aliases = ['awk', 'gawk', 'mawk', 'nawk']
+    filenames = ['*.awk']
+    mimetypes = ['application/x-awk']
+    url = 'https://en.wikipedia.org/wiki/AWK'
+    version_added = '1.5'
+
+    tokens = {
+        'commentsandwhitespace': [
+            (r'\s+', Text),
+            (r'#.*$', Comment.Single)
+        ],
+        'slashstartsregex': [
+            include('commentsandwhitespace'),
+            (r'/(\\.|[^[/\\\n]|\[(\\.|[^\]\\\n])*])+/'
+             r'\B', String.Regex, '#pop'),
+            (r'(?=/)', Text, ('#pop', 'badregex')),
+            default('#pop')
+        ],
+        'badregex': [
+            (r'\n', Text, '#pop')
+        ],
+        'root': [
+            (r'^(?=\s|/)', Text, 'slashstartsregex'),
+            include('commentsandwhitespace'),
+            (r'\+\+|--|\|\||&&|in\b|\$|!?~|\?|:|'
+             r'(\*\*|[-<>+*%\^/!=|])=?', Operator, 'slashstartsregex'),
+            (r'[{(\[;,]', Punctuation, 'slashstartsregex'),
+            (r'[})\].]', Punctuation),
+            (r'(break|continue|do|while|exit|for|if|else|'
+             r'return)\b', Keyword, 'slashstartsregex'),
+            (r'function\b', Keyword.Declaration, 'slashstartsregex'),
+            (r'(atan2|cos|exp|int|log|rand|sin|sqrt|srand|gensub|gsub|index|'
+             r'length|match|split|sprintf|sub|substr|tolower|toupper|close|'
+             r'fflush|getline|next|nextfile|print|printf|strftime|systime|'
+             r'delete|system)\b', Keyword.Reserved),
+            (r'(ARGC|ARGIND|ARGV|BEGIN|CONVFMT|ENVIRON|END|ERRNO|FIELDWIDTHS|'
+             r'FILENAME|FNR|FS|IGNORECASE|NF|NR|OFMT|OFS|ORFS|RLENGTH|RS|'
+             r'RSTART|RT|SUBSEP)\b', Name.Builtin),
+            (r'[$a-zA-Z_]\w*', Name.Other),
+            (r'[0-9][0-9]*\.[0-9]+([eE][0-9]+)?[fd]?', Number.Float),
+            (r'0x[0-9a-fA-F]+', Number.Hex),
+            (r'[0-9]+', Number.Integer),
+            (r'"(\\\\|\\[^\\]|[^"\\])*"', String.Double),
+            (r"'(\\\\|\\[^\\]|[^'\\])*'", String.Single),
+        ]
+    }
+
+
+class SedLexer(RegexLexer):
+    """
+    Lexer for Sed script files.
+    """
+    name = 'Sed'
+    aliases = ['sed', 'gsed', 'ssed']
+    filenames = ['*.sed', '*.[gs]sed']
+    mimetypes = ['text/x-sed']
+    url = 'https://en.wikipedia.org/wiki/Sed'
+    version_added = ''
+    flags = re.MULTILINE
+
+    # Match the contents within delimiters such as /<contents>/
+    _inside_delims = r'((?:(?:\\[^\n]|[^\\])*?\\\n)*?(?:\\.|[^\\])*?)'
+
+    tokens = {
+        'root': [
+            (r'\s+', Whitespace),
+            (r'#.*$', Comment.Single),
+            (r'[0-9]+', Number.Integer),
+            (r'\$', Operator),
+            (r'[{};,!]', Punctuation),
+            (r'[dDFgGhHlnNpPqQxz=]', Keyword),
+            (r'([berRtTvwW:])([^;\n]*)', bygroups(Keyword, String.Single)),
+            (r'([aci])((?:.*?\\\n)*(?:.*?[^\\]$))', bygroups(Keyword, String.Double)),
+            (r'([qQ])([0-9]*)', bygroups(Keyword, Number.Integer)),
+            (r'(/)' + _inside_delims + r'(/)', bygroups(Punctuation, String.Regex, Punctuation)),
+            (r'(\\(.))' + _inside_delims + r'(\2)',
+             bygroups(Punctuation, None, String.Regex, Punctuation)),
+            (r'(y)(.)' + _inside_delims + r'(\2)' + _inside_delims + r'(\2)',
+             bygroups(Keyword, Punctuation, String.Single, Punctuation, String.Single, Punctuation)),
+            (r'(s)(.)' + _inside_delims + r'(\2)' + _inside_delims + r'(\2)((?:[gpeIiMm]|[0-9])*)',
+             bygroups(Keyword, Punctuation, String.Regex, Punctuation, String.Single, Punctuation,
+                      Keyword))
+        ]
+    }
+
+class VimLexer(RegexLexer):
+    """
+    Lexer for VimL script files.
+    """
+    name = 'VimL'
+    aliases = ['vim']
+    filenames = ['*.vim', '.vimrc', '.exrc', '.gvimrc',
+                 '_vimrc', '_exrc', '_gvimrc', 'vimrc', 'gvimrc']
+    mimetypes = ['text/x-vim']
+    url = 'https://www.vim.org'
+    version_added = '0.8'
+
+    flags = re.MULTILINE
+
+    _python = r'py(?:t(?:h(?:o(?:n)?)?)?)?'
+
+    tokens = {
+        'root': [
+            (r'^([ \t:]*)(' + _python + r')([ \t]*)(<<)([ \t]*)(.*)((?:\n|.)*)(\6)',
+             bygroups(using(this), Keyword, Text, Operator, Text, Text,
+                      using(PythonLexer), Text)),
+            (r'^([ \t:]*)(' + _python + r')([ \t])(.*)',
+             bygroups(using(this), Keyword, Text, using(PythonLexer))),
+
+            (r'^\s*".*', Comment),
+
+            (r'[ \t]+', Text),
+            # TODO: regexes can have other delims
+            (r'/[^/\\\n]*(?:\\[\s\S][^/\\\n]*)*/', String.Regex),
+            (r'"[^"\\\n]*(?:\\[\s\S][^"\\\n]*)*"', String.Double),
+            (r"'[^\n']*(?:''[^\n']*)*'", String.Single),
+
+            # Who decided that doublequote was a good comment character??
+            (r'(?<=\s)"[^\-:.%#=*].*', Comment),
+            (r'-?\d+', Number),
+            (r'#[0-9a-f]{6}', Number.Hex),
+            (r'^:', Punctuation),
+            (r'[()<>+=!|,~-]', Punctuation),  # Inexact list.  Looks decent.
+            (r'\b(let|if|else|endif|elseif|fun|function|endfunction)\b',
+             Keyword),
+            (r'\b(NONE|bold|italic|underline|dark|light)\b', Name.Builtin),
+            (r'\b\w+\b', Name.Other),  # These are postprocessed below
+            (r'.', Text),
+        ],
+    }
+
+    def __init__(self, **options):
+        from pygments.lexers._vim_builtins import auto, command, option
+        self._cmd = command
+        self._opt = option
+        self._aut = auto
+
+        RegexLexer.__init__(self, **options)
+
+    def is_in(self, w, mapping):
+        r"""
+        It's kind of difficult to decide if something might be a keyword
+        in VimL because it allows you to abbreviate them.  In fact,
+        'ab[breviate]' is a good example.  :ab, :abbre, or :abbreviate are
+        valid ways to call it so rather than making really awful regexps
+        like::
+
+            \bab(?:b(?:r(?:e(?:v(?:i(?:a(?:t(?:e)?)?)?)?)?)?)?)?\b
+
+        we match `\b\w+\b` and then call is_in() on those tokens.  See
+        `scripts/get_vimkw.py` for how the lists are extracted.
+        """
+        p = bisect(mapping, (w,))
+        if p > 0:
+            if mapping[p-1][0] == w[:len(mapping[p-1][0])] and \
+               mapping[p-1][1][:len(w)] == w:
+                return True
+        if p < len(mapping):
+            return mapping[p][0] == w[:len(mapping[p][0])] and \
+                mapping[p][1][:len(w)] == w
+        return False
+
+    def get_tokens_unprocessed(self, text):
+        # TODO: builtins are only subsequent tokens on lines
+        #       and 'keywords' only happen at the beginning except
+        #       for :au ones
+        for index, token, value in \
+                RegexLexer.get_tokens_unprocessed(self, text):
+            if token is Name.Other:
+                if self.is_in(value, self._cmd):
+                    yield index, Keyword, value
+                elif self.is_in(value, self._opt) or \
+                        self.is_in(value, self._aut):
+                    yield index, Name.Builtin, value
+                else:
+                    yield index, Text, value
+            else:
+                yield index, token, value

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deepseek_v3/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_deepseek_v3 import *
+    from .modeling_deepseek_v3 import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deepseek_v3/configuration_deepseek_v3.py

@@ -0,0 +1,115 @@
+# Copyright 2025 bzantium and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on the DeepSeekV3 implementations from the DeepSeek AI team. (https://huggingface.co/deepseek-ai/DeepSeek-V3)
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""DeepSeekV3 model configuration"""
+
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...modeling_rope_utils import RopeParameters
+from ...utils import auto_docstring
+
+
+@auto_docstring(checkpoint="bzantium/tiny-deepseek-v3")
+@strict
+class DeepseekV3Config(PreTrainedConfig):
+    r"""
+    n_group (`int`, *optional*, defaults to 8):
+        Number of groups for routed experts.
+    first_k_dense_replace (`int`, *optional*, defaults to 3):
+        Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                        \--k dense layers--/
+    rope_interleave (`bool`, *optional*, defaults to `True`):
+        Whether to interleave the rotary position embeddings.
+
+    Example:
+
+    ```python
+    >>> from transformers import DeepseekV3Model, DeepseekV3Config
+
+    >>> # Initializing a Deepseek-V3 style configuration
+    >>> configuration = DeepseekV3Config()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",
+        "layers.*.mlp.experts.down_proj": "rowwise",
+        "layers.*.mlp.experts": "moe_tp_experts",
+        "layers.*.mlp.shared_experts.gate_proj": "colwise",
+        "layers.*.mlp.shared_experts.up_proj": "colwise",
+        "layers.*.mlp.shared_experts.down_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+    attribute_map = {
+        "num_local_experts": "n_routed_experts",
+    }
+
+    vocab_size: int = 129280
+    hidden_size: int = 7168
+    intermediate_size: int = 18432
+    moe_intermediate_size: int = 2048
+    num_hidden_layers: int = 61
+    num_attention_heads: int = 128
+    num_key_value_heads: int | None = 128
+    n_shared_experts: int = 1
+    n_routed_experts: int = 256
+    routed_scaling_factor: float = 2.5
+    kv_lora_rank: int = 512
+    q_lora_rank: int | None = 1536
+    qk_rope_head_dim: int = 64
+    v_head_dim: int | None = 128
+    qk_nope_head_dim: int = 128
+    n_group: int | None = 8
+    topk_group: int | None = 4
+    num_experts_per_tok: int | None = 8
+    first_k_dense_replace: int | None = 3
+    norm_topk_prob: bool | None = True
+    hidden_act: str = "silu"
+    max_position_embeddings: int = 4096
+    initializer_range: float = 0.02
+    rms_norm_eps: float = 1e-6
+    use_cache: bool = True
+    pad_token_id: int | None = None
+    bos_token_id: int | None = 0
+    eos_token_id: int | list[int] | None = 1
+    pretraining_tp: int | None = 1
+    tie_word_embeddings: bool = False
+    rope_parameters: RopeParameters | dict | None = None
+    rope_interleave: bool | None = True
+    attention_bias: bool = False
+    attention_dropout: float | int | None = 0.0
+
+    def __post_init__(self, **kwargs):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+        self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+        self.head_dim = self.qk_rope_head_dim
+        super().__post_init__(**kwargs)
+
+
+__all__ = ["DeepseekV3Config"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deepseek_v3/modeling_deepseek_v3.py

@@ -0,0 +1,722 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/deepseek_v3/modular_deepseek_v3.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_deepseek_v3.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+import math
+from collections.abc import Callable
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ... import initialization as init
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache
+from ...generation import GenerationMixin
+from ...integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernel_func_from_hub
+from ...masking_utils import create_causal_mask
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import (
+    GenericForSequenceClassification,
+    GenericForTokenClassification,
+    GradientCheckpointingLayer,
+)
+from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import is_flash_attention_requested, maybe_autocast, merge_with_config_defaults
+from ...utils.output_capturing import capture_outputs
+from .configuration_deepseek_v3 import DeepseekV3Config
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class DeepseekV3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
+        """
+        DeepseekV3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class DeepseekV3RotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: DeepseekV3Config, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: DeepseekV3Config | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class DeepseekV3MLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class DeepseekV3TopkRouter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.n_routed_experts = config.n_routed_experts
+
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size)))
+        self.register_buffer("e_score_correction_bias", torch.zeros(self.n_routed_experts))
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.view(-1, self.config.hidden_size)
+        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+        return router_logits
+
+
+@use_experts_implementation
+class DeepseekV3NaiveMoe(nn.Module):
+    """Collection of expert weights stored as 3D tensors."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_local_experts
+        self.hidden_dim = config.hidden_size
+        self.intermediate_dim = config.moe_intermediate_size
+        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
+        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        top_k_index: torch.Tensor,
+        top_k_weights: torch.Tensor,
+    ) -> torch.Tensor:
+        final_hidden_states = torch.zeros_like(hidden_states)
+        with torch.no_grad():
+            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
+            expert_mask = expert_mask.permute(2, 1, 0)
+            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+
+        for expert_idx in expert_hit:
+            expert_idx = expert_idx[0]
+            if expert_idx == self.num_experts:
+                continue
+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+            current_state = hidden_states[token_idx]
+            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
+            current_hidden_states = self.act_fn(gate) * up
+            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
+            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
+            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+        return final_hidden_states
+
+
+class DeepseekV3MoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.experts = DeepseekV3NaiveMoe(config)
+        self.gate = DeepseekV3TopkRouter(config)
+        self.shared_experts = DeepseekV3MLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+        self.n_routed_experts = config.n_routed_experts
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.top_k = config.num_experts_per_tok
+
+    def route_tokens_to_experts(self, router_logits):
+        router_logits = router_logits.sigmoid()
+        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
+        group_scores = (
+            router_logits_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .reshape(-1, self.n_routed_experts)
+        )
+        scores_for_choice = router_logits_for_choice.masked_fill(~score_mask.bool(), float("-inf"))
+        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
+        topk_weights = router_logits.gather(1, topk_indices)
+        if self.norm_topk_prob:
+            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weights /= denominator
+        topk_weights = topk_weights * self.routed_scaling_factor
+        return topk_indices, topk_weights
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        router_logits = self.gate(hidden_states)
+        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+@use_kernel_func_from_hub("rotary_pos_emb")
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class DeepseekV3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: DeepseekV3Config, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.q_a_layernorm = DeepseekV3RMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = DeepseekV3RMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        if self.config.rope_parameters.get("rope_type", "default") != "default":
+            mscale_all_dim = self.config.rope_parameters.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_parameters["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scaling = self.scaling * mscale * mscale
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        if self.q_lora_rank is None:
+            q_states = self.q_proj(hidden_states)
+        else:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings
+        if self.config.rope_interleave:  # support using interleaved weights for efficiency
+            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+        else:
+            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
+
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class DeepseekV3DecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: DeepseekV3Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = DeepseekV3Attention(config=config, layer_idx=layer_idx)
+
+        if layer_idx >= config.first_k_dense_replace:
+            self.mlp = DeepseekV3MoE(config)
+        else:
+            self.mlp = DeepseekV3MLP(config)
+
+        self.input_layernorm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+@auto_docstring
+class DeepseekV3PreTrainedModel(PreTrainedModel):
+    config: DeepseekV3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["DeepseekV3DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _can_compile_fullgraph = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": DeepseekV3DecoderLayer,
+        "attentions": DeepseekV3Attention,
+    }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
+    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.61.*"]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, DeepseekV3TopkRouter):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
+        elif isinstance(module, DeepseekV3NaiveMoe):
+            init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
+            init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
+
+
+@auto_docstring
+class DeepseekV3Model(DeepseekV3PreTrainedModel):
+    def __init__(self, config: DeepseekV3Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [DeepseekV3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = DeepseekV3RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @merge_with_config_defaults
+    @capture_outputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if position_ids is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            position_ids = position_ids.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class DeepseekV3ForCausalLM(DeepseekV3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = DeepseekV3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, DeepseekV3ForCausalLM
+
+        >>> model = DeepseekV3ForCausalLM.from_pretrained("meta-deepseek_v3/DeepseekV3-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-deepseek_v3/DeepseekV3-2-7b-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class DeepseekV3ForSequenceClassification(GenericForSequenceClassification, DeepseekV3PreTrainedModel):
+    pass
+
+
+class DeepseekV3ForTokenClassification(GenericForTokenClassification, DeepseekV3PreTrainedModel):
+    pass
+
+
+__all__ = [
+    "DeepseekV3PreTrainedModel",
+    "DeepseekV3Model",
+    "DeepseekV3ForCausalLM",
+    "DeepseekV3ForSequenceClassification",
+    "DeepseekV3ForTokenClassification",
+]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/deepseek_v3/modular_deepseek_v3.py

@@ -0,0 +1,340 @@
+import math
+from collections.abc import Callable
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ... import initialization as init
+from ...cache_utils import Cache
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import GenericForSequenceClassification, GenericForTokenClassification
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import logging
+from ...utils.generic import is_flash_attention_requested
+from ..llama.modeling_llama import (
+    LlamaDecoderLayer,
+    LlamaForCausalLM,
+    LlamaModel,
+    LlamaPreTrainedModel,
+    LlamaRMSNorm,
+    LlamaRotaryEmbedding,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+    rotate_half,
+)
+from ..mixtral.modeling_mixtral import MixtralExperts
+from ..qwen2_moe.modeling_qwen2_moe import Qwen2MoeMLP
+from .configuration_deepseek_v3 import DeepseekV3Config
+
+
+logger = logging.get_logger(__name__)
+
+
+class DeepseekV3RMSNorm(LlamaRMSNorm):
+    pass
+
+
+class DeepseekV3RotaryEmbedding(LlamaRotaryEmbedding):
+    pass
+
+
+class DeepseekV3MLP(Qwen2MoeMLP):
+    pass
+
+
+def apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class DeepseekV3TopkRouter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.n_routed_experts = config.n_routed_experts
+
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size)))
+        self.register_buffer("e_score_correction_bias", torch.zeros(self.n_routed_experts))
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.view(-1, self.config.hidden_size)
+        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+        return router_logits
+
+
+class DeepseekV3NaiveMoe(MixtralExperts):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_experts = config.num_local_experts
+        self.intermediate_dim = config.moe_intermediate_size
+
+
+class DeepseekV3MoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.experts = DeepseekV3NaiveMoe(config)
+        self.gate = DeepseekV3TopkRouter(config)
+        self.shared_experts = DeepseekV3MLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+        self.n_routed_experts = config.n_routed_experts
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.top_k = config.num_experts_per_tok
+
+    def route_tokens_to_experts(self, router_logits):
+        router_logits = router_logits.sigmoid()
+        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
+        group_scores = (
+            router_logits_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .reshape(-1, self.n_routed_experts)
+        )
+        scores_for_choice = router_logits_for_choice.masked_fill(~score_mask.bool(), float("-inf"))
+        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
+        topk_weights = router_logits.gather(1, topk_indices)
+        if self.norm_topk_prob:
+            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weights /= denominator
+        topk_weights = topk_weights * self.routed_scaling_factor
+        return topk_indices, topk_weights
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        router_logits = self.gate(hidden_states)
+        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+class DeepseekV3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: DeepseekV3Config, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.q_a_layernorm = DeepseekV3RMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = DeepseekV3RMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        if self.config.rope_parameters.get("rope_type", "default") != "default":
+            mscale_all_dim = self.config.rope_parameters.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_parameters["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scaling = self.scaling * mscale * mscale
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        if self.q_lora_rank is None:
+            q_states = self.q_proj(hidden_states)
+        else:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings
+        if self.config.rope_interleave:  # support using interleaved weights for efficiency
+            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+        else:
+            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
+
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class DeepseekV3DecoderLayer(LlamaDecoderLayer):
+    def __init__(self, config: DeepseekV3Config, layer_idx: int):
+        nn.Module.__init__(self)
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = DeepseekV3Attention(config=config, layer_idx=layer_idx)
+
+        if layer_idx >= config.first_k_dense_replace:
+            self.mlp = DeepseekV3MoE(config)
+        else:
+            self.mlp = DeepseekV3MLP(config)
+
+        self.input_layernorm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+
+class DeepseekV3PreTrainedModel(LlamaPreTrainedModel):
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
+    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.61.*"]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        PreTrainedModel._init_weights(self, module)
+        if isinstance(module, DeepseekV3TopkRouter):
+            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
+        elif isinstance(module, DeepseekV3NaiveMoe):
+            init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
+            init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
+
+
+class DeepseekV3Model(LlamaModel):
+    pass
+
+
+class DeepseekV3ForCausalLM(LlamaForCausalLM):
+    pass
+
+
+class DeepseekV3ForSequenceClassification(GenericForSequenceClassification, DeepseekV3PreTrainedModel):
+    pass
+
+
+class DeepseekV3ForTokenClassification(GenericForTokenClassification, DeepseekV3PreTrainedModel):
+    pass
+
+
+__all__ = [
+    "DeepseekV3PreTrainedModel",
+    "DeepseekV3Model",
+    "DeepseekV3ForCausalLM",
+    "DeepseekV3ForSequenceClassification",
+    "DeepseekV3ForTokenClassification",
+]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/llava_onevision/__init__.py

@@ -0,0 +1,31 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_llava_onevision import *
+    from .image_processing_llava_onevision import *
+    from .image_processing_pil_llava_onevision import *
+    from .modeling_llava_onevision import *
+    from .processing_llava_onevision import *
+    from .video_processing_llava_onevision import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/llava_onevision/configuration_llava_onevision.py

@@ -0,0 +1,147 @@
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Literal
+
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+from ..auto import CONFIG_MAPPING, AutoConfig
+
+
+@auto_docstring(checkpoint="llava-hf/llava-onevision-qwen2-7b-ov-hf")
+@strict
+class LlavaOnevisionConfig(PreTrainedConfig):
+    r"""
+    image_grid_pinpoints (`List`, *optional*):
+        A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
+        of the form `(height, width)`.
+    vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+        Aspect ratio used when processong image features. The default value is "anyres_max_9".
+
+    Example:
+
+    ```python
+    >>> from transformers import LlavaOnevisionForConditionalGeneration, LlavaOnevisionConfig, SiglipVisionConfig, Qwen2Config
+
+    >>> # Initializing a CLIP-vision config
+    >>> vision_config = SiglipVisionConfig()
+
+    >>> # Initializing a Llama config
+    >>> text_config = Qwen2Config()
+
+    >>> # Initializing a Llava-Next llava-hf/llava-onevision-qwen2-7b-ov-hf style configuration
+    >>> configuration = LlavaOnevisionConfig(vision_config, text_config)
+
+    >>> # Initializing a model from the llava-hf/llava-onevision-qwen2-7b-ov-hf style configuration
+    >>> model = LlavaOnevisionForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "llava_onevision"
+    attribute_map = {
+        "image_token_id": "image_token_index",
+        "video_token_id": "video_token_index",
+    }
+    sub_configs = {"text_config": AutoConfig, "vision_config": AutoConfig}
+
+    vision_config: dict | PreTrainedConfig | None = None
+    text_config: dict | PreTrainedConfig | None = None
+    image_token_index: int = 151646
+    video_token_index: int = 151647
+    projector_hidden_act: str = "gelu"
+    vision_feature_select_strategy: Literal["default", "full"] = "full"
+    vision_feature_layer: int | list[int] = -1
+    multimodal_projector_bias: bool = True
+    tie_word_embeddings: bool = False
+    image_grid_pinpoints: list | None = None
+    vision_aspect_ratio: str = "anyres_max_9"
+
+    def __post_init__(self, **kwargs):
+        if isinstance(self.vision_config, dict):
+            self.vision_config["model_type"] = self.vision_config.get("model_type", "siglip_vision_model")
+            self.vision_config = CONFIG_MAPPING[self.vision_config["model_type"]](**self.vision_config)
+        elif self.vision_config is None:
+            self.vision_config = CONFIG_MAPPING["siglip_vision_model"](
+                hidden_size=1152,
+                intermediate_size=4304,
+                patch_size=14,
+                image_size=384,
+                num_hidden_layers=26,
+                num_attention_heads=16,
+                vision_use_head=False,
+            )
+
+        if isinstance(self.text_config, dict):
+            self.text_config["model_type"] = self.text_config.get("model_type", "qwen2")
+            self.text_config = CONFIG_MAPPING[self.text_config["model_type"]](**self.text_config)
+        elif self.text_config is None:
+            self.text_config = CONFIG_MAPPING["qwen2"]()
+
+        self.image_grid_pinpoints = (
+            self.image_grid_pinpoints
+            if self.image_grid_pinpoints is not None
+            else [
+                [384, 384],
+                [384, 768],
+                [384, 1152],
+                [384, 1536],
+                [384, 1920],
+                [384, 2304],
+                [768, 384],
+                [768, 768],
+                [768, 1152],
+                [768, 1536],
+                [768, 1920],
+                [768, 2304],
+                [1152, 384],
+                [1152, 768],
+                [1152, 1152],
+                [1152, 1536],
+                [1152, 1920],
+                [1152, 2304],
+                [1536, 384],
+                [1536, 768],
+                [1536, 1152],
+                [1536, 1536],
+                [1536, 1920],
+                [1536, 2304],
+                [1920, 384],
+                [1920, 768],
+                [1920, 1152],
+                [1920, 1536],
+                [1920, 1920],
+                [1920, 2304],
+                [2304, 384],
+                [2304, 768],
+                [2304, 1152],
+                [2304, 1536],
+                [2304, 1920],
+                [2304, 2304],
+            ]
+        )
+
+        # The default value is `False` but this config is used with many model types
+        # Attr `tie_word_embeddings` was saved in text config for those models, so we
+        # need an ugly workaround and forward-pass the attr from text config
+        if not self.tie_word_embeddings and self.text_config.tie_word_embeddings:
+            self.tie_word_embeddings = self.text_config.tie_word_embeddings
+
+        super().__post_init__(**kwargs)
+
+
+__all__ = ["LlavaOnevisionConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/llava_onevision/image_processing_pil_llava_onevision.py

@@ -0,0 +1,307 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/llava_onevision/modular_llava_onevision.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_llava_onevision.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import numpy as np
+
+from ...image_processing_backends import PilBackend
+from ...image_processing_utils import BatchFeature, get_patch_output_size, select_best_resolution
+from ...image_transforms import divide_to_patches
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+)
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import TensorType, auto_docstring
+
+
+class LlavaOnevisionImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    image_grid_pinpoints (`list[list[int]]`, *optional*):
+        A list of possible resolutions to use for processing high resolution images. The best resolution is selected
+        based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess`
+        method.
+    """
+
+    image_grid_pinpoints: list[list[int]]
+
+
+@auto_docstring
+class LlavaOnevisionImageProcessorPil(PilBackend):
+    model_input_names = ["pixel_values", "image_sizes", "batch_num_images"]
+    valid_kwargs = LlavaOnevisionImageProcessorKwargs
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 384, "width": 384}
+    default_to_square = False
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    do_pad = True
+    image_grid_pinpoints = [[384, 384], [384, 768], [384, 1152], [384, 1536], [384, 1920], [384, 2304], [768, 384], [768, 768], [768, 1152], [768, 1536], [768, 1920], [768, 2304], [1152, 384], [1152, 768], [1152, 1152], [1152, 1536], [1152, 1920], [1152, 2304], [1536, 384], [1536, 768], [1536, 1152], [1536, 1536], [1536, 1920], [1536, 2304], [1920, 384], [1920, 768], [1920, 1152], [1920, 1536], [1920, 1920], [1920, 2304], [2304, 384], [2304, 768], [2304, 1152], [2304, 1536], [2304, 1920], [2304, 2304]]  # fmt: skip
+
+    def __init__(self, **kwargs: Unpack[LlavaOnevisionImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    @auto_docstring
+    def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionImageProcessorKwargs]) -> BatchFeature:
+        if isinstance(images, (tuple, list)) and isinstance(images[0], (tuple, list)):
+            # if the first element is a list, we assume that all elements are lists
+            images = [x for x in images if x]  # handle text-only case
+            batch_num_images = [len(x) for x in images]
+        elif isinstance(images, (tuple, list)):
+            # treat this as a single-image case for backward compatibility
+            batch_num_images = [1] * len(images)
+        else:
+            batch_num_images = [1]
+        return super().preprocess(images, batch_num_images, **kwargs)
+
+    def _get_padding_size(self, original_resolution: tuple, target_resolution: tuple):
+        """Get padding size for patching (returns tuple format for np.pad)."""
+        original_height, original_width = original_resolution
+        target_height, target_width = target_resolution
+        paste_x, r_x = divmod(target_width - original_width, 2)
+        paste_y, r_y = divmod(target_height - original_height, 2)
+        return (paste_y, paste_y + r_y), (paste_x, paste_x + r_x)
+
+    def _resize_for_patching(
+        self,
+        image: np.ndarray,
+        target_resolution: tuple,
+        resample: PILImageResampling,
+    ) -> np.ndarray:
+        """Resizes an image to a target resolution while maintaining aspect ratio."""
+        new_height, new_width = get_patch_output_size(
+            image, target_resolution, input_data_format=ChannelDimension.FIRST
+        )
+        resized_image = self.resize(image=image, size=SizeDict(height=new_height, width=new_width), resample=resample)
+
+        return resized_image
+
+    def _pad_for_patching(self, image: np.ndarray, target_resolution: tuple) -> np.ndarray:
+        """Pad an image to a target resolution while maintaining aspect ratio."""
+        new_resolution = get_patch_output_size(image, target_resolution, input_data_format=ChannelDimension.FIRST)
+        padding_hw = self._get_padding_size(new_resolution, target_resolution)
+
+        # For channels_first format (C, H, W), add (0, 0) for channel dimension
+        # padding_hw is ((before_h, after_h), (before_w, after_w))
+        # np.pad expects ((before_C, after_C), (before_H, after_H), (before_W, after_W))
+        padding = ((0, 0), padding_hw[0], padding_hw[1])
+
+        # Use np.pad directly for patching padding
+        padded_image = np.pad(image, padding, mode="constant", constant_values=0)
+
+        return padded_image
+
+    def get_image_patches(
+        self,
+        image: np.ndarray,
+        grid_pinpoints: list[list[int]],
+        size: tuple,
+        patch_size: int,
+        resample: PILImageResampling,
+    ) -> list[np.ndarray]:
+        """Process an image with variable resolutions by dividing it into patches."""
+        if not isinstance(grid_pinpoints, list):
+            raise TypeError("grid_pinpoints must be a list of possible resolutions.")
+
+        possible_resolutions = grid_pinpoints
+
+        image_size = image.shape[-2:]
+        best_resolution = select_best_resolution(image_size, possible_resolutions)
+        resized_image = self._resize_for_patching(image, best_resolution, resample=resample)
+        padded_image = self._pad_for_patching(resized_image, best_resolution)
+
+        patches = divide_to_patches(padded_image, patch_size=patch_size)
+
+        size_height, size_width = size
+        resized_original_image = self.resize(
+            image=image,
+            size=SizeDict(height=size_height, width=size_width),
+            resample=resample,
+        )
+
+        image_patches = [resized_original_image] + patches
+
+        return image_patches
+
+    def _pad_for_batching(
+        self,
+        pixel_values: list[np.ndarray],
+    ) -> list[np.ndarray]:
+        """Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches."""
+        max_patch = max(len(x) for x in pixel_values)
+        # Use np.pad directly for patch dimension padding
+        padded_values = []
+        for image in pixel_values:
+            # Padding format: ((before_dim0, after_dim0), (before_dim1, after_dim1), ...)
+            padding = ((0, max_patch - image.shape[0]), (0, 0), (0, 0), (0, 0))
+            padded_image = np.pad(image, padding, mode="constant", constant_values=0)
+            padded_values.append(padded_image)
+
+        return padded_values
+
+    def _preprocess(
+        self,
+        images: list[np.ndarray],
+        batch_num_images: list[int],
+        do_resize: bool,
+        size: SizeDict,
+        image_grid_pinpoints: list[list[int]],
+        resample: "PILImageResampling | None",
+        do_center_crop: bool,
+        crop_size: SizeDict,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: float | list[float] | None,
+        image_std: float | list[float] | None,
+        do_pad: bool | None,
+        pad_size: SizeDict | None,
+        disable_grouping: bool | None,
+        return_tensors: str | TensorType | None,
+        **kwargs,
+    ) -> BatchFeature:
+        """Custom preprocessing for LLaVA-NeXT with patch processing."""
+        processed_images = []
+        image_sizes = []
+        # only single image patching is supported
+        need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+
+        # Backend's resize method handles resample conversion, so we can pass it directly
+        # Determine the size tuple
+        if size and size.height and size.width:
+            size_tuple = (size.height, size.width)
+        else:
+            size_tuple = (size.shortest_edge, size.shortest_edge)
+
+        # Determine the patch size
+        if crop_size and crop_size.height:
+            patch_size = crop_size.height
+        elif size and size.height:
+            patch_size = size.height
+        else:
+            patch_size = size.shortest_edge
+
+        for i, image in enumerate(images):
+            if need_patching[i]:
+                # convert image into a list of patches
+                # we intentionally use the same data format as the input data format
+                image_patches = self.get_image_patches(
+                    image,
+                    image_grid_pinpoints,
+                    size=size_tuple,
+                    patch_size=patch_size,
+                    resample=resample,
+                )
+            else:
+                padded_image = self.pad_to_square(
+                    image=image, background_color=tuple(int(x * 255) for x in self.image_mean)
+                )
+                image_patches = [padded_image]
+
+            # preprocess patches
+            pixel_values = []
+            for patch in image_patches:
+                if do_resize:
+                    patch = self.resize(image=patch, size=size, resample=resample)
+
+                if do_center_crop:
+                    patch = self.center_crop(image=patch, size=crop_size)
+
+                if do_rescale:
+                    patch = self.rescale(image=patch, scale=rescale_factor)
+
+                if do_normalize:
+                    patch = self.normalize(image=patch, mean=image_mean, std=image_std)
+
+                pixel_values.append(patch)
+
+            pixel_values = np.array(pixel_values)
+            processed_images.append(pixel_values)
+            image_sizes.append(image.shape[-2:])
+
+        if do_pad:
+            processed_images = self._pad_for_batching(processed_images)
+
+        return BatchFeature(
+            data={
+                "pixel_values": processed_images,
+                "image_sizes": image_sizes,
+                "batch_num_images": batch_num_images,
+            },
+            tensor_type=return_tensors,
+        )
+
+    def pad_to_square(
+        self,
+        image: np.ndarray,
+        background_color: int | tuple[int, int, int] = 0,
+    ) -> np.ndarray:
+        """
+        Pads an image to a square based on the longest edge.
+
+        Args:
+            image (`np.ndarray`):
+                The image to pad. Shape: (num_channels, height, width) - always channels_first in backend.
+            background_color (`int` or `tuple[int, int, int]`, *optional*, defaults to 0):
+                The color to use for the padding.
+
+        Returns:
+            `np.ndarray`: The padded image.
+        """
+        # Backend always uses channels_first format: (num_channels, height, width)
+        num_channels, height, width = image.shape
+
+        if height == width:
+            return image
+
+        max_dim = max(height, width)
+
+        # Ensure background_color is the correct shape
+        if isinstance(background_color, int):
+            background_color = [background_color]
+        elif len(background_color) != num_channels:
+            raise ValueError(
+                f"background_color must have no more than {num_channels} elements to match the number of channels"
+            )
+
+        result = np.zeros((num_channels, max_dim, max_dim), dtype=image.dtype)
+        for i, color in enumerate(background_color):
+            result[i, :, :] = color
+        if width > height:
+            start = (max_dim - height) // 2
+            result[:, start : start + height, :] = image
+        else:
+            start = (max_dim - width) // 2
+            result[:, :, start : start + width] = image
+
+        return result
+
+
+__all__ = ["LlavaOnevisionImageProcessorPil"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/llava_onevision/modeling_llava_onevision.py

@@ -0,0 +1,848 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/llava_onevision/modular_llava_onevision.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_llava_onevision.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+from torch import nn
+
+from ... import initialization as init
+from ...activations import ACT2FN
+from ...cache_utils import Cache
+from ...generation import GenerationMixin
+from ...image_processing_utils import select_best_resolution
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, ModelOutput
+from ...modeling_utils import PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import TransformersKwargs, auto_docstring, torch_compilable_check
+from ...utils.generic import can_return_tuple, merge_with_config_defaults
+from ..auto import AutoModel
+from .configuration_llava_onevision import LlavaOnevisionConfig
+
+
+@auto_docstring(
+    custom_intro="""
+    Base class for Llava outputs, with hidden states and attentions.
+    """
+)
+@dataclass
+class LlavaOnevisionModelOutputWithPast(BaseModelOutputWithPast):
+    r"""
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    image_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
+        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    video_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor`  of size `(batch_size * num_frames, num_videos, sequence_length, hidden_size)`.
+        video_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    """
+
+    image_hidden_states: torch.FloatTensor | None = None
+
+    video_hidden_states: torch.FloatTensor | None = None
+
+
+@auto_docstring(
+    custom_intro="""
+    Base class for LlavaOnevision causal language model (or autoregressive) outputs.
+    """
+)
+@dataclass
+class LlavaOnevisionCausalLMOutputWithPast(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+        Language modeling loss (for next-token prediction).
+    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    image_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor` of size (batch_size * num_patches, num_images, sequence_length, hidden_size)`.
+        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    video_hidden_states (`torch.FloatTensor`, *optional*):
+        A `torch.FloatTensor`  of size `(batch_size * num_frames, num_videos, sequence_length, hidden_size)`.
+        video_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
+    """
+
+    loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor | None = None
+    past_key_values: Cache | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+    attentions: tuple[torch.FloatTensor] | None = None
+    image_hidden_states: torch.FloatTensor | None = None
+
+    video_hidden_states: torch.FloatTensor | None = None
+
+
+@auto_docstring
+class LlavaOnevisionPreTrainedModel(PreTrainedModel):
+    config: LlavaOnevisionConfig
+    base_model_prefix = "model"
+    input_modalities = ("image", "video", "text")
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+
+    _supports_flash_attn = True
+    _supports_sdpa = True
+
+    _can_compile_fullgraph = True
+    _supports_flex_attn = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": "LlamaDecoderLayer",
+        "attentions": "LlamaAttention",
+    }
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, LlavaOnevisionModel):
+            embed_std = 1 / math.sqrt(self.config.text_config.hidden_size)
+            init.normal_(module.image_newline, mean=0.0, std=embed_std)
+
+
+class LlavaOnevisionMultiModalProjector(nn.Module):
+    def __init__(self, config: LlavaOnevisionConfig):
+        super().__init__()
+        # We have hidden_size * the number of vision feature layers
+        num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size * num_feature_layers,
+            config.text_config.hidden_size,
+            bias=config.multimodal_projector_bias,
+        )
+        self.act = ACT2FN[config.projector_hidden_act]
+        self.linear_2 = nn.Linear(
+            config.text_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
+        )
+
+    def forward(self, image_features):
+        hidden_states = self.linear_1(image_features)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
+    """
+    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
+
+    Args:
+        image_size (`tuple`):
+            The size of the input image in the format (width, height).
+        grid_pinpoints (`List`):
+            A list containing possible resolutions. Each item in the list should be a tuple or list
+            of the form `(height, width)`.
+        patch_size (`int`):
+            The size of each image patch.
+
+    Returns:
+        tuple: The shape of the image patch grid in the format (width, height).
+    """
+    if not isinstance(grid_pinpoints, list):
+        raise TypeError("grid_pinpoints should be a list of tuples or lists")
+
+    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
+    if not isinstance(image_size, (list, tuple)):
+        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
+            raise TypeError(
+                f"image_size invalid type: {type(image_size)} not valid, should be either list, tuple, np.ndarray or tensor"
+            )
+        image_size = image_size.tolist()
+
+    height, width = select_best_resolution(image_size, grid_pinpoints)
+    return height // patch_size, width // patch_size
+
+
+def image_size_to_num_patches(image_size, grid_pinpoints, patch_size: int):
+    """
+    Calculate the number of patches after the preprocessing for images of any resolution.
+
+    Args:
+        image_size (`torch.LongTensor` or `np.ndarray` or `tuple[int, int]`):
+            The size of the input image in the format (height, width). ?
+        grid_pinpoints (`List`):
+            A list containing possible resolutions. Each item in the list should be a tuple or list
+            of the form `(height, width)`.
+        patch_size (`int`):
+            The size of each image patch.
+
+    Returns:
+        int: the number of patches
+    """
+    if not isinstance(grid_pinpoints, list):
+        raise TypeError("grid_pinpoints should be a list of tuples or lists")
+
+    # ! VERY IMPORTANT if image_size is tensor, must convert to into tuple, otherwise it will cause wrong calculate
+    if not isinstance(image_size, (list, tuple)):
+        if not isinstance(image_size, (torch.Tensor, np.ndarray)):
+            raise TypeError(f"image_size invalid type {type(image_size)} with value {image_size}")
+        image_size = image_size.tolist()
+
+    best_resolution = select_best_resolution(image_size, grid_pinpoints)
+    height, width = best_resolution
+    num_patches = 0
+    # consider change to ceil(height/patch_size)*ceil(width/patch_size) + 1
+    for i in range(0, height, patch_size):
+        for j in range(0, width, patch_size):
+            num_patches += 1
+    # add the base patch
+    num_patches += 1
+    return num_patches
+
+
+def unpad_image(tensor, original_size):
+    """
+    Unpads a PyTorch tensor of a padded and resized image.
+
+    Args:
+        tensor (`torch.Tensor`):
+            The image tensor, assumed to be of shape (num_channels, height, width).
+        original_size (`tuple`):
+            The original size of the image (height, width).
+
+    Returns:
+        `torch.Tensor`: The unpadded image tensor.
+    """
+    if not isinstance(original_size, (list, tuple)):
+        if not isinstance(original_size, (torch.Tensor, np.ndarray)):
+            raise TypeError(
+                f"image_size invalid type: {type(original_size)} not valid, should be either list, tuple, np.ndarray or tensor"
+            )
+        original_size = original_size.tolist()
+    original_height, original_width = original_size
+    current_height, current_width = tensor.shape[1:]
+
+    original_aspect_ratio = original_width / original_height
+    current_aspect_ratio = current_width / current_height
+
+    if original_aspect_ratio > current_aspect_ratio:
+        scale_factor = current_width / original_width
+        new_height = int(round(original_height * scale_factor, 7))
+        padding = (current_height - new_height) // 2
+        unpadded_tensor = tensor[:, padding : current_height - padding, :]
+    else:
+        scale_factor = current_height / original_height
+        new_width = int(round(original_width * scale_factor, 7))
+        padding = (current_width - new_width) // 2
+        unpadded_tensor = tensor[:, :, padding : current_width - padding]
+
+    return unpadded_tensor
+
+
+@auto_docstring(
+    custom_intro="""
+    The Llava-Next model which consists of a vision backbone and a language model without language modeling head.
+    """
+)
+class LlavaOnevisionModel(LlavaOnevisionPreTrainedModel):
+    base_model_prefix = "model"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.vision_tower = AutoModel.from_config(config.vision_config)
+
+        self.multi_modal_projector = LlavaOnevisionMultiModalProjector(config)
+        embed_std = 1 / math.sqrt(config.text_config.hidden_size)
+        self.image_newline = nn.Parameter(torch.randn(config.text_config.hidden_size, dtype=self.dtype) * embed_std)
+
+        self.vocab_size = config.text_config.vocab_size
+        self.language_model = AutoModel.from_config(config.text_config)
+        self.post_init()
+
+    def pack_image_features(self, image_features, image_sizes, image_newline=None, vision_aspect_ratio="anyres_max_9"):
+        """
+        Reshape, unpad and then pack each image_feature into a single image_features tensor containing all visual vectors.
+
+        Args:
+            image_features (`list[torch.Tensor]` of length num_images, each of shape `(num_patches, image_length, embed_dim)`)
+                List of image feature tensor, each contains all the visual feature of all patches.
+            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
+                Actual image size of each images (H, W).
+            image_newline (`torch.Tensor` of shape `(embed_dim)`)
+                New line embedding vector.
+            vision_aspect_ratio (`str`, *optional*, "anyres_max_9"):
+                Aspect ratio used when processing image features. The default value is "anyres_max_9".
+        Returns:
+            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
+            feature_lens (`list[int]`)
+                token length of each image in image_features
+        """
+        new_image_features = []
+        feature_lens = []
+        for image_idx, image_feature in enumerate(image_features):
+            if image_feature.shape[0] > 1:
+                base_image_feature = image_feature[0]
+                image_feature = image_feature[1:]
+                height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
+                if height * width != base_image_feature.shape[0]:
+                    raise ValueError("The number of patches is not consistent with the image size.")
+                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
+                    image_sizes[image_idx],
+                    self.config.image_grid_pinpoints,
+                    self.config.vision_config.image_size,
+                )
+                image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
+                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                max_num_patches = int(vision_aspect_ratio.strip("anyres_max_"))
+                channels, curr_height, curr_width = image_feature.shape
+                ratio = math.sqrt(curr_height * curr_width / (max_num_patches * height**2))
+                if ratio > 1.1:
+                    image_feature = image_feature[None]
+                    image_feature = nn.functional.interpolate(
+                        image_feature, [int(curr_height // ratio), int(curr_width // ratio)], mode="bilinear"
+                    )[0]
+                if image_newline is not None:
+                    image_feature = torch.cat(
+                        (
+                            image_feature,
+                            image_newline[:, None, None]
+                            .expand(*image_feature.shape[:-1], 1)
+                            .to(image_feature.device, image_feature.dtype),
+                        ),
+                        dim=-1,
+                    )
+                image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+                image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+            else:
+                image_feature = image_feature[0]
+                if image_newline is not None:
+                    image_feature = torch.cat((image_feature, image_newline[None].to(image_feature)), dim=0)
+            new_image_features.append(image_feature)
+            feature_lens.append(image_feature.size(0))
+        feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features[0].device)
+        return new_image_features, feature_lens
+
+    @merge_with_config_defaults
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
+    )
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_sizes: torch.Tensor,
+        vision_feature_layer: int | list[int] | list[int] | None = None,
+        vision_feature_select_strategy: str | None = None,
+        vision_aspect_ratio: str | None = None,
+        batch_num_images: torch.LongTensor | None = None,
+        output_hidden_states: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        image_sizes (`torch.Tensor` of shape `(num_images, 2)`):
+            Actual image size of each images (H, W).
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processing image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        """
+        # ! infer image_num_patches from image_sizes
+        if batch_num_images is None:
+            # treat this as a single-image case for backward compatibility
+            need_patching = [True] * len(image_sizes)
+        else:
+            need_patching = [n == 1 for n in batch_num_images for _ in range(n)]
+        image_num_patches = [
+            image_size_to_num_patches(
+                image_size=imsize,
+                grid_pinpoints=self.config.image_grid_pinpoints,
+                patch_size=self.config.vision_config.image_size,
+            )
+            if should_patch
+            else 1
+            for imsize, should_patch in zip(image_sizes, need_patching)
+        ]
+        if pixel_values.dim() == 5:
+            # stacked if input is (batch_size, num_patches, num_channels, height, width)
+            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
+            pixel_values = torch.cat(_pixel_values_list, dim=0)
+        elif pixel_values.dim() != 4:
+            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
+            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
+
+        image_outputs = self.vision_tower(
+            pixel_values,
+            output_hidden_states=True,  # Ignore arg on purpose
+            return_dict=True,
+            **kwargs,
+        )
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_image_feature = torch.cat(hs_pool, dim=-1)
+
+        if vision_feature_select_strategy == "default":
+            selected_image_feature = selected_image_feature[:, 1:]
+        image_features = self.multi_modal_projector(selected_image_feature)
+        image_features = torch.split(image_features, image_num_patches, dim=0)
+
+        image_features, feature_lens = self.pack_image_features(
+            image_features,
+            image_sizes,
+            image_newline=self.image_newline,
+            vision_aspect_ratio=vision_aspect_ratio,
+        )
+        image_outputs.pooler_output = image_features
+
+        return image_outputs
+
+    def get_placeholder_mask(
+        self,
+        input_ids: torch.LongTensor,
+        inputs_embeds: torch.FloatTensor,
+        image_features: torch.FloatTensor | None = None,
+        video_features: torch.FloatTensor | None = None,
+    ):
+        """
+        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
+        equal to the length of multimodal features. If the lengths are different, an error is raised.
+        """
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
+            )
+            special_image_mask = special_image_mask.all(-1)
+            special_video_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
+            )
+            special_video_mask = special_video_mask.all(-1)
+        else:
+            special_image_mask = input_ids == self.config.image_token_id
+            special_video_mask = input_ids == self.config.video_token_id
+
+        n_image_tokens = special_image_mask.sum()
+        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+        if image_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_image_mask].numel() == image_features.numel(),
+                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}",
+            )
+
+        n_video_tokens = special_video_mask.sum()
+        special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+        if video_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_video_mask].numel() == video_features.numel(),
+                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
+            )
+        return special_image_mask, special_video_mask
+
+    @merge_with_config_defaults
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        pixel_values: torch.FloatTensor | None = None,
+        image_sizes: torch.LongTensor | None = None,
+        pixel_values_videos: torch.FloatTensor | None = None,
+        image_sizes_videos: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        vision_feature_layer: int | list[int] | list[int] | None = None,
+        vision_feature_select_strategy: str | None = None,
+        vision_aspect_ratio: str | None = None,
+        batch_num_images: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple | LlavaOnevisionModelOutputWithPast:
+        r"""
+        image_sizes_videos (`torch.LongTensor` of shape `(batch_size, frames, 2)`, *optional*):
+            The sizes of the videos in the batch, being (height, width) for each frame in the video.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processing image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        """
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        image_features = None
+        if pixel_values is not None:
+            image_features = self.get_image_features(
+                pixel_values,
+                image_sizes,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+                batch_num_images=batch_num_images,
+                return_dict=True,
+            ).pooler_output
+            image_features = torch.cat(image_features, dim=0)
+            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+            special_image_mask, _ = self.get_placeholder_mask(
+                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
+            )
+            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+        video_features = None
+        if pixel_values_videos is not None:
+            video_features = self.get_video_features(
+                pixel_values_videos,
+                vision_feature_layer=vision_feature_layer,
+                vision_feature_select_strategy=vision_feature_select_strategy,
+                return_dict=True,
+            ).pooler_output
+            image_newline = (
+                self.image_newline[None, None, :].repeat(video_features.shape[0], 1, 1).to(video_features.device)
+            )
+            video_features = torch.cat((video_features, image_newline), dim=1)
+            video_features = video_features.flatten(0, 1).to(inputs_embeds.device, inputs_embeds.dtype)
+            _, special_video_mask = self.get_placeholder_mask(
+                input_ids, inputs_embeds=inputs_embeds, video_features=video_features
+            )
+            inputs_embeds = inputs_embeds.masked_scatter(special_video_mask, video_features)
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            **kwargs,
+        )
+
+        return LlavaOnevisionModelOutputWithPast(
+            last_hidden_state=outputs.last_hidden_state,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=image_features,
+            video_hidden_states=video_features,
+        )
+
+    @merge_with_config_defaults
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="Obtains video last hidden states from the vision tower, apply multimodal projection and pooling."
+    )
+    def get_video_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: int | list[int] | list[int] | None = None,
+        vision_feature_select_strategy: str | None = None,
+        output_hidden_states: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
+            The tensors corresponding to the input video.
+        vision_feature_layer (`Union[int, list[int]], *optional*, defaults to -2`):
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
+        vision_feature_select_strategy (`str`):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`
+        """
+        batch_size, frames, channels, height, width = pixel_values.shape
+        pixel_values = pixel_values.view(batch_size * frames, channels, height, width)
+        vision_outputs = self.vision_tower(
+            pixel_values,
+            output_hidden_states=True,  # Ignore arg on purpose
+            return_dict=True,
+            **kwargs,
+        )
+
+        # If we have one vision feature layer, return the corresponding hidden states,
+        # otherwise, select the hidden states of each feature layer and concatenate them
+        if isinstance(vision_feature_layer, int):
+            selected_video_feature = vision_outputs.hidden_states[vision_feature_layer]
+        else:
+            hs_pool = [vision_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
+            selected_video_feature = torch.cat(hs_pool, dim=-1)
+
+        if vision_feature_select_strategy == "default":
+            selected_video_feature = selected_video_feature[:, 1:]
+        video_features = self.multi_modal_projector(selected_video_feature)
+
+        video_features = self.apply_pooling(video_features)
+        video_features = video_features.reshape(batch_size, frames * video_features.shape[1], -1)
+        vision_outputs.pooler_output = video_features
+
+        return vision_outputs
+
+    def apply_pooling(self, image_features):
+        height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size
+        batch_frames, seq_len, dim = image_features.shape
+        image_features = image_features.view(batch_frames, height, width, -1)
+        image_features = image_features.permute(0, 3, 1, 2).contiguous()
+
+        height, width = image_features.shape[2:]
+        scaled_shape = [math.ceil(height / 2), math.ceil(width / 2)]
+        image_features = nn.functional.interpolate(image_features, size=scaled_shape, mode="bilinear")
+
+        image_features = image_features.permute(0, 2, 3, 1)
+        image_features = image_features.view(batch_frames, -1, dim)
+        return image_features
+
+
+@auto_docstring(
+    custom_intro="""
+    The LLAVA-NeXT model which consists of a vision backbone and a language model.
+    """
+)
+class LlavaOnevisionForConditionalGeneration(LlavaOnevisionPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
+
+    def __init__(self, config: LlavaOnevisionConfig):
+        super().__init__(config)
+        self.model = LlavaOnevisionModel(config)
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.lm_head
+
+    def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
+        return self.model.pack_image_features(
+            image_features=image_features,
+            image_sizes=image_sizes,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            image_newline=image_newline,
+        )
+
+    @merge_with_config_defaults
+    @can_return_tuple
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_sizes: torch.Tensor,
+        vision_feature_layer: int | list[int] | list[int] | None = None,
+        vision_feature_select_strategy: str | None = None,
+        vision_aspect_ratio: str | None = None,
+        batch_num_images: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        image_sizes (`torch.Tensor` of shape `(num_images, 2)`):
+            Actual image size of each images (H, W).
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processing image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        """
+        return self.model.get_image_features(
+            pixel_values=pixel_values,
+            image_sizes=image_sizes,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            vision_aspect_ratio=vision_aspect_ratio,
+            batch_num_images=batch_num_images,
+            **kwargs,
+        )
+
+    @merge_with_config_defaults
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        pixel_values: torch.FloatTensor | None = None,
+        image_sizes: torch.LongTensor | None = None,
+        pixel_values_videos: torch.FloatTensor | None = None,
+        image_sizes_videos: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        vision_feature_layer: int | list[int] | list[int] | None = None,
+        vision_feature_select_strategy: str | None = None,
+        vision_aspect_ratio: str | None = None,
+        batch_num_images: torch.LongTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | LlavaOnevisionCausalLMOutputWithPast:
+        r"""
+        image_sizes_videos (`torch.LongTensor` of shape `(batch_size, frames, 2)`, *optional*):
+            The sizes of the videos in the batch, being (height, width) for each frame in the video.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processing image features. The default value is "anyres_max_9".
+        batch_num_images (`torch.LongTensor`, *optional*):
+            Number of images in each sample.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import httpx
+        >>> from io import BytesIO
+        >>> import torch
+        >>> from transformers import LlavaOnevisionProcessor, LlavaOnevisionForConditionalGeneration
+
+        >>> model = LlavaOnevisionForConditionalGeneration.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", dtype="float16", device_map="cuda:0")
+        >>> processor = LlavaOnevisionProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
+
+        >>> conversation = [
+        ...     {
+        ...       "role": "user",
+        ...       "content": [
+        ...           {"type": "text", "text": "What is shown in this image?"},
+        ...           {"type": "image"},
+        ...         ],
+        ...     },
+        ... ]
+        >>> prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
+        >>> inputs = processor(text=prompt, images=image, return_tensors='pt').to(0, torch.float16)
+
+        >>> output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        >>> processor.batch_decode(output, skip_special_tokens=True)[0]
+        "user\n\nWhat is shown in this image?\nassistant\ncat"
+        ```"""
+        outputs = self.model(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            pixel_values_videos=pixel_values_videos,
+            image_sizes=image_sizes,
+            image_sizes_videos=image_sizes_videos,
+            vision_aspect_ratio=vision_aspect_ratio,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            batch_num_images=batch_num_images,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
+            )
+
+        return LlavaOnevisionCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            image_hidden_states=outputs.image_hidden_states,
+            video_hidden_states=outputs.video_hidden_states,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        image_sizes=None,
+        pixel_values_videos=None,
+        image_sizes_videos=None,
+        attention_mask=None,
+        logits_to_keep=None,
+        is_first_iteration=False,
+        **kwargs,
+    ):
+        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            logits_to_keep=logits_to_keep,
+            is_first_iteration=is_first_iteration,
+            **kwargs,
+        )
+
+        if is_first_iteration or not kwargs.get("use_cache", True):
+            # Pixel values are used only in the first iteration if available
+            # In subsequent iterations, they are already merged with text and cached
+            # NOTE: first iteration doesn't have to be prefill, it can be the first
+            # iteration with a question and cached system prompt (continue generate from cache)
+            model_inputs["pixel_values"] = pixel_values
+            model_inputs["image_sizes"] = image_sizes
+            model_inputs["pixel_values_videos"] = pixel_values_videos
+            model_inputs["image_sizes_videos"] = image_sizes_videos
+
+        return model_inputs
+
+    @merge_with_config_defaults
+    @can_return_tuple
+    @auto_docstring
+    def get_video_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        vision_feature_layer: int | list[int] | list[int] | None = None,
+        vision_feature_select_strategy: str | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_frames, channels, height, width)`)
+            The tensors corresponding to the input video.
+        vision_feature_layer (`Union[int, list[int]]`, *optional;*):
+            The index of the layer to select the vision feature. If multiple indices are provided,
+            the vision feature of the corresponding indices will be concatenated to form the
+            vision features.
+        vision_feature_select_strategy (`str`, *optional*):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`
+        """
+        return self.model.get_video_features(
+            pixel_values=pixel_values,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            **kwargs,
+        )
+
+
+__all__ = ["LlavaOnevisionModel", "LlavaOnevisionForConditionalGeneration", "LlavaOnevisionPreTrainedModel"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/llava_onevision/processing_llava_onevision.py

@@ -0,0 +1,290 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for LLaVa-Onevision.
+"""
+
+import math
+from collections.abc import Iterable
+
+import numpy as np
+
+from ...feature_extraction_utils import BatchFeature
+from ...image_processing_utils import select_best_resolution
+from ...image_utils import ImageInput, get_image_size, to_numpy_array
+from ...processing_utils import MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack
+from ...tokenization_utils_base import PreTokenizedInput, TextInput
+from ...utils import auto_docstring, logging
+from ...video_utils import VideoInput
+
+
+logger = logging.get_logger(__name__)
+
+
+class LlavaOnevisionProcessorKwargs(ProcessingKwargs, total=False):
+    # see processing_utils.ProcessingKwargs documentation for usage.
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+            "return_mm_token_type_ids": False,
+        },
+        "image_kwargs": {},
+    }
+
+
+@auto_docstring
+class LlavaOnevisionProcessor(ProcessorMixin):
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        video_processor=None,
+        num_image_tokens=None,
+        vision_feature_select_strategy=None,
+        chat_template=None,
+        image_token="<image>",
+        video_token="<video>",
+        vision_aspect_ratio="anyres_max_9",
+        **kwargs,
+    ):
+        r"""
+        num_image_tokens (`int`, *optional*):
+            Number of image tokens for one imagethat will be returned by vision tower.
+        vision_feature_select_strategy (`str`, *optional*):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Should be same as in model's config
+        image_token (`str`, *optional*, defaults to `"<image>"`):
+            Special token used to denote image location.
+        video_token (`str`, *optional*, defaults to `"<video>"`):
+            Special token used to denote video location.
+        vision_aspect_ratio (`str`, *optional*, defaults to `"anyres_max_9"`):
+            Aspect ratio used when processong image features. The default value is "anyres_max_9".
+        """
+        self.num_image_tokens = num_image_tokens
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.image_token = tokenizer.image_token if hasattr(tokenizer, "image_token") else image_token
+        self.video_token = tokenizer.video_token if hasattr(tokenizer, "video_token") else video_token
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.image_token)
+        )
+        self.video_token_id = (
+            tokenizer.video_token_id
+            if getattr(tokenizer, "video_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.video_token)
+        )
+        self.vision_aspect_ratio = vision_aspect_ratio
+        super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
+
+    @auto_docstring
+    def __call__(
+        self,
+        images: ImageInput | None = None,
+        text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None,
+        videos: VideoInput | None = None,
+        **kwargs: Unpack[LlavaOnevisionProcessorKwargs],
+    ) -> BatchFeature:
+        r"""
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **pixel_values_videos** -- Pixel values of a video input to be fed to a model. Returned when `videos` is not `None`.
+            - **image_sizes** -- Size of each image that will be used to unpad an image. Returned when `images` is not `None`.
+        """
+
+        output_kwargs = self._merge_kwargs(
+            LlavaOnevisionProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise TypeError("Invalid input text. Please provide a string, or a list of strings")
+
+        image_inputs = video_inputs = {}
+
+        if images is not None:
+            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
+
+            batch_num_images = iter(image_inputs["batch_num_images"])
+            image_sizes = iter(image_inputs["image_sizes"])
+            height, width = get_image_size(
+                to_numpy_array(image_inputs["pixel_values"][0][0]),
+                channel_dim=output_kwargs["images_kwargs"].get("data_format"),
+            )
+            text, num_image_tokens = self._expand_image_tokens(
+                text, image_sizes, height, width, self.image_token, batch_num_images
+            )
+
+        if videos is not None:
+            video_inputs = self.video_processor(videos, **output_kwargs["videos_kwargs"])
+
+            one_video = video_inputs.get("pixel_values_videos")[0]
+            if isinstance(video_inputs.get("pixel_values_videos")[0], (list, tuple)):
+                one_video = np.array(one_video)
+            else:
+                one_video = to_numpy_array(one_video)
+            height, width = get_image_size(one_video[0], channel_dim=output_kwargs["images_kwargs"].get("data_format"))
+            num_frames = one_video.shape[0]  # frame dim is always after batch dim
+            patches_height_width = int(math.sqrt(self.num_image_tokens))
+            pooled_height_width = math.ceil(patches_height_width / 2)
+            num_video_tokens = (num_frames * pooled_height_width * pooled_height_width) + 1  # +1 for newline token
+            text = [sample.replace(self.video_token, self.video_token * num_video_tokens) for sample in text]
+
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
+        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        self._check_special_mm_tokens(text, text_inputs, modalities=["image"])
+
+        if return_mm_token_type_ids:
+            text_inputs["mm_token_type_ids"] = self.create_mm_token_type_ids(text_inputs["input_ids"])
+        return BatchFeature(data={**text_inputs, **image_inputs, **video_inputs}, tensor_type=return_tensors)
+
+    def _expand_image_tokens(
+        self,
+        text: list[TextInput],
+        image_sizes: Iterable[list[int] | int],
+        height: int,
+        width: int,
+        special_token: str,
+        batch_num_images: Iterable[int],
+    ):
+        prompt_strings = []
+        max_num_vision_tokens = 0
+        for sample in text:
+            if special_token in sample:
+                num_images = next(batch_num_images)  # should consume iterable
+                is_multi_image = num_images != 1
+            else:
+                is_multi_image = False
+            while special_token in sample:
+                original_size = next(image_sizes)  # should consume iterable
+                if is_multi_image:
+                    num_image_tokens = self.num_image_tokens + 1  # one for image_newline
+                else:
+                    if not isinstance(original_size, (list, tuple)):
+                        # cast to list to avoid numerical precision errors when calculating unpadding
+                        original_size = original_size.tolist()
+                    orig_height, orig_width = original_size
+                    num_image_tokens = self._get_number_of_features(orig_height, orig_width, height, width)
+                max_num_vision_tokens = max(max_num_vision_tokens, num_image_tokens)
+                if self.vision_feature_select_strategy == "default":
+                    num_image_tokens -= 1
+                sample = sample.replace(special_token, "<placeholder>" * num_image_tokens, 1)
+            prompt_strings.append(sample)
+        text = [sample.replace("<placeholder>", special_token) for sample in prompt_strings]
+        return text, max_num_vision_tokens
+
+    def _get_number_of_features(self, orig_height: int, orig_width: int, height: int, width: int) -> int:
+        image_grid_pinpoints = self.image_processor.image_grid_pinpoints
+
+        height_best_resolution, width_best_resolution = select_best_resolution(
+            [orig_height, orig_width], image_grid_pinpoints
+        )
+        scale_height, scale_width = height_best_resolution // height, width_best_resolution // width
+
+        patches_height = patches_width = int(math.sqrt(self.num_image_tokens))
+        unpadded_features, newline_features = self._get_unpadded_features(
+            orig_height, orig_width, patches_height, patches_width, scale_height, scale_width
+        )
+
+        # The base patch covers the entire image (no CLS for SigLIP)
+        base_features = self.num_image_tokens
+        num_image_tokens = unpadded_features + newline_features + base_features
+        return num_image_tokens
+
+    # Adapted from transformers.models.llava_next.processing_llava_next.LlavaNextProcessor._get_unpadded_features
+    def _get_unpadded_features(self, height, width, patches_height, patches_width, scale_height, scale_width):
+        """
+        Get number of features for a given image with height/width. LLaVA-NeXT is different from LLaVA
+        because it divided each image into patches depending on its resolution. Therefore we need to calculate how many
+        patches an image is divided into and get the number of features from that.
+        """
+        current_height = patches_height * scale_height
+        current_width = patches_width * scale_width
+
+        original_aspect_ratio = width / height
+        current_aspect_ratio = current_width / current_height
+        if original_aspect_ratio > current_aspect_ratio:
+            new_height = int(round(height * (current_width / width), 7))
+            padding = (current_height - new_height) // 2
+            current_height -= padding * 2
+        else:
+            new_width = int(round(width * (current_height / height), 7))
+            padding = (current_width - new_width) // 2
+            current_width -= padding * 2
+
+        unpadded_features = current_height * current_width
+        newline_features = current_height
+
+        max_num_patches = int(self.vision_aspect_ratio.strip("anyres_max_"))
+        ratio = math.sqrt(current_height * current_width / (max_num_patches * patches_height**2))
+        if ratio > 1.1:
+            unpadded_features = int(current_height // ratio) * int(current_width // ratio)
+            newline_features = int(current_height // ratio)
+
+        return (unpadded_features, newline_features)
+
+    def _get_num_multimodal_tokens(self, image_sizes=None, video_sizes=None, **kwargs):
+        """
+        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
+        Args:
+            image_sizes (list[list[str]], *optional*):
+                The input sizes formatted as (height, width) per each image.
+            video_sizes (list[list[str]], *optional*):
+                The input sizes formatted as (num_frames, height, width) per each video.
+            audio_lengths (list[int], *optional*):
+                The input length formatted as per each audio.
+        Returns:
+            dict[str, list[int]]: A dictionary mapping each modality ("image", "video", "audio")
+            to a list containing the number of placeholder tokens required. If the model doesn't accept
+            a certain modality or no input sizes are provided, the dict value is set to an empty list.
+        """
+        vision_data = {}
+        if image_sizes is not None:
+            images_kwargs = LlavaOnevisionProcessorKwargs._defaults.get("images_kwargs", {})
+            images_kwargs.update(kwargs)
+
+            size = images_kwargs.get("size", None) or self.image_processor.size
+            size = (
+                (size["shortest_edge"], size["shortest_edge"])
+                if "shortest_edge" in size
+                else (min(size["height"], size["width"]), min(size["height"], size["width"]))
+            )
+            processed_height, processed_width = size
+
+            batch_num_image_tokens = []
+            num_image_patches = [1] * len(image_sizes)  # llava-ov doesn't batch pixels as Idefics, thus `1` patch`
+            for image_size in image_sizes:
+                orig_height, orig_width = image_size
+                num_image_tokens = self._get_number_of_features(
+                    orig_height, orig_width, processed_height, processed_width
+                )
+                if self.vision_feature_select_strategy == "default":
+                    num_image_tokens -= 1
+                batch_num_image_tokens.append(num_image_tokens)
+            vision_data.update({"num_image_tokens": batch_num_image_tokens, "num_image_patches": num_image_patches})
+
+        return MultiModalData(**vision_data)
+
+
+__all__ = ["LlavaOnevisionProcessor"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/llava_onevision/video_processing_llava_onevision.py

@@ -0,0 +1,36 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Video processor class for LLaVa-Onevision."""
+
+from ...image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling
+from ...video_processing_utils import BaseVideoProcessor
+
+
+class LlavaOnevisionVideoProcessor(BaseVideoProcessor):
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 384, "width": 384}
+    rescale_factor = 1 / 255
+    default_to_square = False
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    do_sample_frames = False  # Set to False for BC, recommended to set `True` in new models
+
+
+__all__ = ["LlavaOnevisionVideoProcessor"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/pegasus/configuration_pegasus.py

@@ -0,0 +1,76 @@
+# Copyright 2021, Google and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PEGASUS model configuration"""
+
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+
+
+@auto_docstring(checkpoint="google/pegasus-large")
+@strict
+class PegasusConfig(PreTrainedConfig):
+    r"""
+    Example:
+
+    ```python
+    >>> from transformers import PegasusConfig, PegasusModel
+
+    >>> # Initializing a PEGASUS google/pegasus-large style configuration
+    >>> configuration = PegasusConfig()
+
+    >>> # Initializing a model (with random weights) from the google/pegasus-large style configuration
+    >>> model = PegasusModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "pegasus"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_attention_heads": "encoder_attention_heads",
+        "hidden_size": "d_model",
+        "num_hidden_layers": "encoder_layers",
+    }
+
+    vocab_size: int = 50265
+    max_position_embeddings: int = 1024
+    encoder_layers: int = 12
+    encoder_ffn_dim: int = 4096
+    encoder_attention_heads: int = 16
+    decoder_layers: int = 12
+    decoder_ffn_dim: int = 4096
+    decoder_attention_heads: int = 16
+    encoder_layerdrop: float | int = 0.0
+    decoder_layerdrop: float | int = 0.0
+    use_cache: bool = True
+    is_encoder_decoder: bool = True
+    activation_function: str = "gelu"
+    d_model: int = 1024
+    dropout: float | int = 0.1
+    attention_dropout: float | int = 0.0
+    activation_dropout: float | int = 0.0
+    init_std: float = 0.02
+    decoder_start_token_id: int | None = 0
+    scale_embedding: bool = False
+    pad_token_id: int | None = 0
+    eos_token_id: int | list[int] | None = 1
+    forced_eos_token_id: int | list[int] | None = 1
+    is_decoder: bool = False
+    tie_word_embeddings: bool = True
+
+
+__all__ = ["PegasusConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/pegasus/modeling_pegasus.py

@@ -0,0 +1,1132 @@
+# Copyright 2021, Google and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch PEGASUS model."""
+
+import copy
+import math
+from collections.abc import Callable
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ... import initialization as init
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
+from ...generation import GenerationMixin
+from ...masking_utils import create_bidirectional_mask, create_causal_mask
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import (
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_torchdynamo_compiling,
+    logging,
+)
+from ...utils.generic import merge_with_config_defaults
+from ...utils.output_capturing import OutputRecorder, capture_outputs
+from .configuration_pegasus import PegasusConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("self.model.config.pad_token_id has to be defined.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.marian.modeling_marian.MarianSinusoidalPositionalEmbedding with Marian->Pegasus
+class PegasusSinusoidalPositionalEmbedding(nn.Embedding):
+    """This module produces sinusoidal positional embeddings of any length."""
+
+    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: int | None = None) -> None:
+        super().__init__(num_positions, embedding_dim, _freeze=True)
+
+    def create_weight(self):
+        """
+        Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
+        the 2nd half of the vector. [dim // 2:]
+        """
+        n_pos, dim = self.weight.shape
+        position_enc = np.array(
+            [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
+        )
+        out = torch.empty(n_pos, dim, dtype=self.weight.dtype, requires_grad=False)
+        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
+        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+        return out
+
+    @torch.no_grad()
+    def forward(
+        self, input_ids_shape: torch.Size, past_key_values_length: int = 0, position_ids: torch.Tensor | None = None
+    ) -> torch.Tensor:
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        if position_ids is None:
+            bsz, seq_len = input_ids_shape[:2]
+            position_ids = torch.arange(
+                past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
+            )
+        return super().forward(position_ids)
+
+
+# Copied from transformers.models.bert.modeling_bert.eager_attention_forward
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float | None = None,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    if scaling is None:
+        scaling = query.size(-1) ** -0.5
+
+    # Take the dot product between "query" and "key" to get the raw attention scores.
+    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
+
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+
+    attn_output = torch.matmul(attn_weights, value)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Pegasus
+class PegasusAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+        is_causal: bool = False,
+        config: PegasusConfig | None = None,
+        layer_idx: int | None = None,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        self.config = config
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+        self.is_causal = is_causal
+        self.layer_idx = layer_idx
+        if layer_idx is None and self.is_decoder:
+            logger.warning_once(
+                f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
+                "will lead to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: torch.Tensor | None = None,
+        past_key_values: Cache | None = None,
+        attention_mask: torch.Tensor | None = None,
+        # TODO: we need a refactor so that the different attention modules can get their specific kwargs
+        # ATM, we have mixed things encoder, decoder, and encoder-decoder attn
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        # determine input shapes
+        input_shape = hidden_states.shape[:-1]
+
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        # get query proj
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        is_updated = False
+        if past_key_values is not None:
+            if isinstance(past_key_values, EncoderDecoderCache):
+                is_updated = past_key_values.is_updated.get(self.layer_idx)
+                if is_cross_attention:
+                    # after the first generated id, we can subsequently re-use all key/value_states from cache
+                    curr_past_key_values = past_key_values.cross_attention_cache
+                else:
+                    curr_past_key_values = past_key_values.self_attention_cache
+            else:
+                curr_past_key_values = past_key_values
+
+        current_states = key_value_states if is_cross_attention else hidden_states
+        if is_cross_attention and past_key_values is not None and is_updated:
+            # reuse k,v, cross_attentions
+            key_states = curr_past_key_values.layers[self.layer_idx].keys
+            value_states = curr_past_key_values.layers[self.layer_idx].values
+        else:
+            key_states = self.k_proj(current_states)
+            value_states = self.v_proj(current_states)
+            kv_shape = (*current_states.shape[:-1], -1, self.head_dim)
+            key_states = key_states.view(kv_shape).transpose(1, 2)
+            value_states = value_states.view(kv_shape).transpose(1, 2)
+
+            if past_key_values is not None:
+                key_states, value_states = curr_past_key_values.update(key_states, value_states, self.layer_idx)
+                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
+                if is_cross_attention and isinstance(past_key_values, EncoderDecoderCache):
+                    past_key_values.is_updated[self.layer_idx] = True
+
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Pegasus, MBART->PEGASUS
+class PegasusEncoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: PegasusConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = PegasusAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+            config=config,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            **kwargs,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        if hidden_states.dtype == torch.float16:
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        return hidden_states
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Pegasus, MBART->PEGASUS
+class PegasusDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: PegasusConfig, layer_idx: int | None = None):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = PegasusAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+            is_causal=True,
+            config=config,
+            layer_idx=layer_idx,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = PegasusAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+            config=config,
+            layer_idx=layer_idx,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        encoder_hidden_states: torch.Tensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = True,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            past_key_values (`Cache`): cached past key and value projection states
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            **kwargs,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            hidden_states, _ = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                past_key_values=past_key_values,
+                **kwargs,
+            )
+            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+@auto_docstring
+class PegasusPreTrainedModel(PreTrainedModel):
+    config: PegasusConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = True
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, PegasusSinusoidalPositionalEmbedding):
+            init.copy_(module.weight, module.create_weight())
+        elif isinstance(module, PegasusForConditionalGeneration):
+            init.zeros_(module.final_logits_bias)
+
+
+class PegasusEncoder(PegasusPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`PegasusEncoderLayer`].
+
+    Args:
+        config: PegasusConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    _can_record_outputs = {
+        "hidden_states": PegasusEncoderLayer,
+        "attentions": PegasusAttention,
+    }
+
+    def __init__(self, config: PegasusConfig):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = PegasusSinusoidalPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([PegasusEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def resize_position_embeddings(self, new_num_position_embeddings: int):
+        """
+        Resizes position embeddings matrix of the model if `new_num_position_embeddings !=
+        config.max_position_embeddings`.
+
+        Arguments:
+            new_num_position_embeddings (`int`):
+                The number of new position embeddings. If position embeddings are learned, increasing the size will add
+                newly initialized vectors at the end, whereas reducing the size will remove vectors from the end. If
+                position embeddings are not learned (*e.g.* sinusoidal position embeddings), increasing the size will
+                add correct vectors at the end following the position encoding algorithm, whereas reducing the size
+                will remove vectors from the end.
+        """
+        logger.info(f"Setting `config.max_position_embeddings={new_num_position_embeddings}`...")
+        self.config.max_position_embeddings = new_num_position_embeddings
+
+        self.embed_positions = PegasusSinusoidalPositionalEmbedding(
+            self.config.max_position_embeddings,
+            self.config.d_model,
+            self.padding_idx,
+        )
+        init.copy_(self.embed_positions.weight, self.embed_positions.create_weight())
+        self.embed_positions.to(self.device)
+
+    def get_position_embeddings(self) -> nn.Embedding:
+        """
+        Returns the position embeddings matrix
+        """
+        return self.embed_positions
+
+    @merge_with_config_defaults
+    @capture_outputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutput:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        input_shape = inputs_embeds.shape[:-1]
+        embed_pos = self.embed_positions(input_shape)
+
+        hidden_states = inputs_embeds + embed_pos
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        attention_mask = create_bidirectional_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+        )
+
+        for idx, encoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
+            to_drop = False
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:  # skip the layer
+                    to_drop = True
+
+            if not to_drop:
+                hidden_states = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    **kwargs,
+                )
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+        )
+
+
+class PegasusDecoder(PegasusPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`PegasusDecoderLayer`]
+
+    Args:
+        config: PegasusConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    _can_record_outputs = {
+        "hidden_states": PegasusDecoderLayer,
+        "attentions": OutputRecorder(PegasusAttention, index=1, layer_name="self_attn"),
+        "cross_attentions": OutputRecorder(PegasusAttention, index=1, layer_name="encoder_attn"),
+    }
+
+    def __init__(self, config: PegasusConfig):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = PegasusSinusoidalPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+            self.padding_idx,
+        )
+        self.layers = nn.ModuleList([PegasusDecoderLayer(config, layer_idx=i) for i in range(config.decoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def resize_position_embeddings(self, new_num_position_embeddings: int):
+        """
+        Resizes position embeddings matrix of the model if `new_num_position_embeddings !=
+        config.max_position_embeddings`.
+
+        Arguments:
+            new_num_position_embeddings (`int`):
+                The number of new position embeddings. If position embeddings are learned, increasing the size will add
+                newly initialized vectors at the end, whereas reducing the size will remove vectors from the end. If
+                position embeddings are not learned (*e.g.* sinusoidal position embeddings), increasing the size will
+                add correct vectors at the end following the position encoding algorithm, whereas reducing the size
+                will remove vectors from the end.
+        """
+        logger.info(f"Setting `config.max_position_embeddings={new_num_position_embeddings}`...")
+        self.config.max_position_embeddings = new_num_position_embeddings
+
+        self.embed_positions = PegasusSinusoidalPositionalEmbedding(
+            self.config.max_position_embeddings,
+            self.config.d_model,
+            self.padding_idx,
+        )
+        init.copy_(self.embed_positions.weight, self.embed_positions.create_weight())
+        self.embed_positions.to(self.device)
+
+    def get_position_embeddings(self) -> nn.Embedding:
+        """
+        Returns the position embeddings matrix
+        """
+        return self.embed_positions
+
+    @merge_with_config_defaults
+    @capture_outputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPastAndCrossAttentions:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        # important to apply scale outside of `if` in case users pass `embeds`
+        inputs_embeds = inputs_embeds * self.embed_scale
+
+        # initialize `past_key_values`
+        if use_cache and past_key_values is None:
+            past_key_values = (
+                EncoderDecoderCache(DynamicCache(config=self.config), DynamicCache(config=self.config))
+                if encoder_hidden_states is not None or self.config.is_encoder_decoder
+                else DynamicCache(config=self.config)
+            )
+
+        batch_size, seq_length = inputs_embeds.size()[:-1]
+        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
+        position_ids = torch.arange(seq_length, device=inputs_embeds.device) + past_key_values_length
+
+        if attention_mask is None and not is_torchdynamo_compiling():
+            # required mask seq length can be calculated via length of past cache
+            mask_seq_length = past_key_values_length + seq_length
+            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
+
+        self_attn_cache = (
+            past_key_values.self_attention_cache
+            if isinstance(past_key_values, EncoderDecoderCache)
+            else past_key_values
+        )
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=self_attn_cache,
+        )
+        encoder_attention_mask = create_bidirectional_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=encoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+
+        # embed positions
+        positions = self.embed_positions((batch_size, seq_length), past_key_values_length, position_ids=position_ids)
+        hidden_states = inputs_embeds + positions
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:
+                    continue
+
+            hidden_states = decoder_layer(
+                hidden_states,
+                causal_mask,
+                encoder_hidden_states,  # as a positional argument for gradient checkpointing
+                encoder_attention_mask=encoder_attention_mask,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                **kwargs,
+            )
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class PegasusModel(PegasusPreTrainedModel):
+    _tied_weights_keys = {
+        "decoder.embed_tokens.weight": "shared.weight",
+        "encoder.embed_tokens.weight": "shared.weight",
+    }
+
+    def __init__(self, config: PegasusConfig):
+        super().__init__(config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.encoder = PegasusEncoder(config)
+        self.decoder = PegasusDecoder(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, value):
+        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+    def resize_position_embeddings(self, new_num_position_embeddings: int):
+        """
+        Resizes position embeddings matrix of the model if `new_num_position_embeddings !=
+        config.max_position_embeddings`.
+
+        Arguments:
+            new_num_position_embeddings (`int`):
+                The number of new position embeddings. If position embeddings are learned, increasing the size will add
+                newly initialized vectors at the end, whereas reducing the size will remove vectors from the end. If
+                position embeddings are not learned (*e.g.* sinusoidal position embeddings), increasing the size will
+                add correct vectors at the end following the position encoding algorithm, whereas reducing the size
+                will remove vectors from the end.
+        """
+        self.config.max_position_embeddings = new_num_position_embeddings
+        self.encoder.resize_position_embeddings(new_num_position_embeddings)
+        self.decoder.resize_position_embeddings(new_num_position_embeddings)
+
+    def get_position_embeddings(self) -> tuple[nn.Embedding]:
+        """
+        Returns the position embeddings matrix
+        """
+        return (self.encoder.get_position_embeddings(), self.decoder.get_position_embeddings())
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        decoder_input_ids: torch.Tensor | None = None,
+        decoder_attention_mask: torch.Tensor | None = None,
+        encoder_outputs: tuple[torch.FloatTensor] | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        decoder_inputs_embeds: torch.Tensor | None = None,
+        use_cache: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | Seq2SeqModelOutput:
+        r"""
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            Pegasus uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, PegasusModel
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("google/pegasus-large")
+        >>> model = PegasusModel.from_pretrained("google/pegasus-large")
+
+        >>> inputs = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt")
+        >>> decoder_inputs = tokenizer("Studies show that", return_tensors="pt")
+        >>> outputs = model(input_ids=inputs.input_ids, decoder_input_ids=decoder_inputs.input_ids)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+        >>> list(last_hidden_states.shape)
+        [1, 4, 1024]
+        ```"""
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                **kwargs,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput
+        elif not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_values, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            **kwargs,
+        )
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@auto_docstring(
+    custom_intro="""
+    The PEGASUS Model with a language modeling head. Can be used for summarization.
+    """
+)
+class PegasusForConditionalGeneration(PegasusPreTrainedModel, GenerationMixin):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = ["final_logits_bias"]
+    _tied_weights_keys = {
+        "lm_head.weight": "model.shared.weight",
+    }
+
+    def __init__(self, config: PegasusConfig):
+        super().__init__(config)
+        self.model = PegasusModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def resize_token_embeddings(
+        self, new_num_tokens: int, pad_to_multiple_of: int | None = None, mean_resizing: bool = True
+    ) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens, pad_to_multiple_of, mean_resizing)
+        self._resize_final_logits_bias(new_embeddings.weight.shape[0])
+        return new_embeddings
+
+    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
+        old_num_tokens = self.final_logits_bias.shape[-1]
+        if new_num_tokens <= old_num_tokens:
+            new_bias = self.final_logits_bias[:, :new_num_tokens]
+        else:
+            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
+            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
+        self.register_buffer("final_logits_bias", new_bias)
+
+    def resize_position_embeddings(self, new_num_position_embeddings: int):
+        """
+        Resizes position embeddings matrix of the model if `new_num_position_embeddings !=
+        config.max_position_embeddings`.
+
+        Arguments:
+            new_num_position_embeddings (`int`):
+                The number of new position embeddings. If position embeddings are learned, increasing the size will add
+                newly initialized vectors at the end, whereas reducing the size will remove vectors from the end. If
+                position embeddings are not learned (*e.g.* sinusoidal position embeddings), increasing the size will
+                add correct vectors at the end following the position encoding algorithm, whereas reducing the size
+                will remove vectors from the end.
+        """
+        self.config.max_position_embeddings = new_num_position_embeddings
+        self.model.encoder.resize_position_embeddings(new_num_position_embeddings)
+        self.model.decoder.resize_position_embeddings(new_num_position_embeddings)
+
+    def get_position_embeddings(self) -> tuple[nn.Embedding]:
+        """
+        Returns the position embeddings matrix
+        """
+        return (self.model.encoder.get_position_embeddings(), self.model.decoder.get_position_embeddings())
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        decoder_input_ids: torch.Tensor | None = None,
+        decoder_attention_mask: torch.Tensor | None = None,
+        encoder_outputs: tuple[torch.FloatTensor] | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        decoder_inputs_embeds: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+        use_cache: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | Seq2SeqLMOutput:
+        r"""
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            Pegasus uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example Summarization:
+
+        ```python
+        >>> from transformers import AutoTokenizer, PegasusForConditionalGeneration
+
+        >>> model = PegasusForConditionalGeneration.from_pretrained("google/pegasus-xsum")
+        >>> tokenizer = AutoTokenizer.from_pretrained("google/pegasus-xsum")
+
+        >>> ARTICLE_TO_SUMMARIZE = (
+        ...     "PG&E stated it scheduled the blackouts in response to forecasts for high winds "
+        ...     "amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were "
+        ...     "scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."
+        ... )
+        >>> inputs = tokenizer(ARTICLE_TO_SUMMARIZE, max_length=1024, return_tensors="pt")
+
+        >>> # Generate Summary
+        >>> summary_ids = model.generate(inputs["input_ids"])
+        >>> tokenizer.batch_decode(summary_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "California's largest electricity provider has turned off power to hundreds of thousands of customers."
+        ```
+        """
+        if labels is not None:
+            if use_cache:
+                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
+            use_cache = False
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs: Seq2SeqModelOutput = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            **kwargs,
+        )
+        lm_logits = self.lm_head(outputs.last_hidden_state) + self.final_logits_bias
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        return Seq2SeqLMOutput(
+            loss=masked_lm_loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
+        return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartDecoderWrapper with Bart->Pegasus
+class PegasusDecoderWrapper(PegasusPreTrainedModel):
+    """
+    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
+    used in combination with the [`EncoderDecoderModel`] framework.
+    """
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.decoder = PegasusDecoder(config)
+        self.post_init()
+
+    def forward(self, *args, **kwargs):
+        return self.decoder(*args, **kwargs)
+
+
+class PegasusForCausalLM(PegasusPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {
+        "lm_head.weight": "model.decoder.embed_tokens.weight",
+    }
+
+    def __init__(self, config):
+        config = copy.deepcopy(config)
+        config.is_decoder = True
+        config.is_encoder_decoder = False
+        super().__init__(config)
+        self.model = PegasusDecoderWrapper(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+    def get_position_embeddings(self) -> nn.Embedding:
+        """
+        Returns the position embeddings matrix
+        """
+        return self.model.decoder.get_position_embeddings()
+
+    def resize_position_embeddings(self, new_num_position_embeddings: int):
+        """
+        Resizes position embeddings matrix of the model if `new_num_position_embeddings !=
+        config.max_position_embeddings`.
+
+        Arguments:
+            new_num_position_embeddings (`int`):
+                The number of new position embeddings. If position embeddings are learned, increasing the size will add
+                newly initialized vectors at the end, whereas reducing the size will remove vectors from the end. If
+                position embeddings are not learned (*e.g.* sinusoidal position embeddings), increasing the size will
+                add correct vectors at the end following the position encoding algorithm, whereas reducing the size
+                will remove vectors from the end.
+        """
+        self.config.max_position_embeddings = new_num_position_embeddings
+        self.model.decoder.resize_position_embeddings(new_num_position_embeddings)
+
+    @can_return_tuple
+    @auto_docstring
+    # Copied from transformers.models.bart.modeling_bart.BartForCausalLM.forward with Bart->Pegasus, facebook/bart-base->google/pegasus-large
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        encoder_hidden_states: torch.FloatTensor | None = None,
+        encoder_attention_mask: torch.FloatTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | CausalLMOutputWithCrossAttentions:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, PegasusForCausalLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("google/pegasus-large")
+        >>> model = PegasusForCausalLM.from_pretrained("google/pegasus-large")
+        >>> assert model.config.is_decoder, f"{model.__class__} has to be configured as a decoder."
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> logits = outputs.logits
+        >>> expected_shape = [1, inputs.input_ids.shape[-1], model.config.vocab_size]
+        >>> list(logits.shape) == expected_shape
+        True
+        ```"""
+
+        outputs: BaseModelOutputWithPastAndCrossAttentions = self.model.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+__all__ = ["PegasusForCausalLM", "PegasusForConditionalGeneration", "PegasusModel", "PegasusPreTrainedModel"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/wavlm/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_wavlm import *
+    from .modeling_wavlm import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)