Add files using upload-large-folder tool

LTA_openwebtext_dualt/logs/lm1b_len128_lognormal_atoms_sde_watch/infer_lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522_step_0037000_logistic_normal_t1p45.log

@@ -0,0 +1,76 @@
+[watch-lognormal-sde] 2026-05-23_02:01:03 infer runs/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0037000.pt -> docs/lta_samples/metrics_20260522/lm1b_len128_lognormal_atoms_every1k_logistic_normal_sde_t1p45_steps128_n256/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0037000
+[load] runs/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0037000.pt
+[ckpt] step=37000
+[sde] generated 16/256
+[sde] generated 32/256
+[sde] generated 48/256
+[sde] generated 64/256
+[sde] generated 80/256
+[sde] generated 96/256
+[sde] generated 112/256
+[sde] generated 128/256
+[sde] generated 144/256
+[sde] generated 160/256
+[sde] generated 176/256
+[sde] generated 192/256
+[sde] generated 208/256
+[sde] generated 224/256
+[sde] generated 240/256
+[sde] generated 256/256
+[score] loading scorer: /e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-large-standard
+[summary] {
+  "type": "summary",
+  "checkpoint": "runs/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0037000.pt",
+  "step": 37000,
+  "decode": {
+    "decode_rule": "logistic_normal_resample_sde",
+    "steps": 128,
+    "model_t_mode": "const0.5",
+    "mean_mode": "anchor_semantic",
+    "endpoint_floor": 0.0,
+    "concentration_min": 1.0,
+    "concentration_max": 1024.0,
+    "endpoint_temp": 1.45,
+    "support_power": 1.0,
+    "semantic_power": 1.0,
+    "noise_init": "logistic_normal",
+    "noise_sigma": 3.0,
+    "noise_dirichlet_concentration": 1.0,
+    "sde_resample": "logistic_normal",
+    "logistic_normal_sigma_min": 0.18,
+    "logistic_normal_sigma_max": 3.0,
+    "logistic_normal_tau_min": 0.65,
+    "logistic_normal_tau_max": 1.0,
+    "final_from": "blend_0.5",
+    "n_samples": 256,
+    "seed": 20260522
+  },
+  "raw_genppl": {
+    "ppl": 37.067606022965414,
+    "nll_per_token": 3.6127434351734498,
+    "tokens": 30524,
+    "kept_samples": 256,
+    "total_samples": 256,
+    "empty_rate": 0.0,
+    "skipped_samples": 0
+  },
+  "stripped_genppl": {
+    "ppl": 48.49552325374488,
+    "nll_per_token": 3.8814714896366596,
+    "tokens": 25687,
+    "kept_samples": 256,
+    "total_samples": 256,
+    "empty_rate": 0.0,
+    "skipped_samples": 0
+  },
+  "diversity": {
+    "sample_entropy": 3.1397080459572364,
+    "unique_tokens": 1860,
+    "token_count": 32768,
+    "distinct_1": 0.0567626953125,
+    "distinct_2": 0.2785740649606299,
+    "top_token_mass": 0.2530517578125
+  }
+}
+[done] docs/lta_samples/metrics_20260522/lm1b_len128_lognormal_atoms_every1k_logistic_normal_sde_t1p45_steps128_n256/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0037000/sde_steps128_samples256_scored.jsonl
+[watch-lognormal-sde] 2026-05-23_02:02:31 done step_0037000

LTA_openwebtext_dualt/logs/lm1b_len128_lognormal_atoms_sde_watch/infer_lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522_step_0055000_logistic_normal_t1p45.log

@@ -0,0 +1,76 @@
+[watch-lognormal-sde] 2026-05-23_03:41:01 infer runs/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0055000.pt -> docs/lta_samples/metrics_20260522/lm1b_len128_lognormal_atoms_every1k_logistic_normal_sde_t1p45_steps128_n256/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0055000
+[load] runs/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0055000.pt
+[ckpt] step=55000
+[sde] generated 16/256
+[sde] generated 32/256
+[sde] generated 48/256
+[sde] generated 64/256
+[sde] generated 80/256
+[sde] generated 96/256
+[sde] generated 112/256
+[sde] generated 128/256
+[sde] generated 144/256
+[sde] generated 160/256
+[sde] generated 176/256
+[sde] generated 192/256
+[sde] generated 208/256
+[sde] generated 224/256
+[sde] generated 240/256
+[sde] generated 256/256
+[score] loading scorer: /e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-large-standard
+[summary] {
+  "type": "summary",
+  "checkpoint": "runs/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0055000.pt",
+  "step": 55000,
+  "decode": {
+    "decode_rule": "logistic_normal_resample_sde",
+    "steps": 128,
+    "model_t_mode": "const0.5",
+    "mean_mode": "anchor_semantic",
+    "endpoint_floor": 0.0,
+    "concentration_min": 1.0,
+    "concentration_max": 1024.0,
+    "endpoint_temp": 1.45,
+    "support_power": 1.0,
+    "semantic_power": 1.0,
+    "noise_init": "logistic_normal",
+    "noise_sigma": 3.0,
+    "noise_dirichlet_concentration": 1.0,
+    "sde_resample": "logistic_normal",
+    "logistic_normal_sigma_min": 0.18,
+    "logistic_normal_sigma_max": 3.0,
+    "logistic_normal_tau_min": 0.65,
+    "logistic_normal_tau_max": 1.0,
+    "final_from": "blend_0.5",
+    "n_samples": 256,
+    "seed": 20260522
+  },
+  "raw_genppl": {
+    "ppl": 20.48864099827913,
+    "nll_per_token": 3.0198706349306113,
+    "tokens": 31075,
+    "kept_samples": 256,
+    "total_samples": 256,
+    "empty_rate": 0.0,
+    "skipped_samples": 0
+  },
+  "stripped_genppl": {
+    "ppl": 20.39658888593573,
+    "nll_per_token": 3.015367675395035,
+    "tokens": 27814,
+    "kept_samples": 256,
+    "total_samples": 256,
+    "empty_rate": 0.0,
+    "skipped_samples": 0
+  },
+  "diversity": {
+    "sample_entropy": 2.5694837759949234,
+    "unique_tokens": 1321,
+    "token_count": 32768,
+    "distinct_1": 0.040313720703125,
+    "distinct_2": 0.19725024606299213,
+    "top_token_mass": 0.27593994140625
+  }
+}
+[done] docs/lta_samples/metrics_20260522/lm1b_len128_lognormal_atoms_every1k_logistic_normal_sde_t1p45_steps128_n256/lta_lm1b_classic_c1024_len128_lognormalatoms_gbs512_4gpu_20260522/step_0055000/sde_steps128_samples256_scored.jsonl
+[watch-lognormal-sde] 2026-05-23_03:42:28 done step_0055000

LTA_openwebtext_dualt/mini_owt_fit/logs/mini_owt_fit_bert_len1024_C1_to_1024_absrope_time4_d768_l12_h12_full_gbs512_8gpu_20260526_151837.log

@@ -0,0 +1,48 @@
+W0526 15:18:38.686000 10232 torch/distributed/run.py:792] 
+W0526 15:18:38.686000 10232 torch/distributed/run.py:792] *****************************************
+W0526 15:18:38.686000 10232 torch/distributed/run.py:792] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0526 15:18:38.686000 10232 torch/distributed/run.py:792] *****************************************
+skipping len=623 text='WHAT?!??! I know. That’s what you’re saying right '
+skipping len=767 text='A notorious protester convicted of wilfully promot'
+skipping len=476 text='× Some Seattle businesses closed for ‘A Day Withou'
+skipping len=610 text='Today, Toyota announced changes in executives’ are'
+skipping len=844 text='North Korean leader Kim Jong Un. AP Images / Busin'
+skipping len=1015 text='We’ve always pictured Scandinavia as the home of g'
+skipping len=177 text="Story highlights Tyka Nelson says her brother's fa"
+skipping len=250 text='There’s measuring the drapes, and then there’s mea'
+skipping len=530 text='Attention! This news was published on the old vers'
+skipping len=591 text='Ad blockers are often painted as the enemy of onli'
+skipping len=362 text='Get cool in-game extras with amiibo accessories! J'
+skipping len=254 text='Stanley “Boom” Williams decided to enter the 2017 '
+skipping len=566 text='About This Game Casino Blackjack 21 with a TWIST!!'
+skipping len=589 text='F ancy cars have always been an important element '
+skipping len=386 text='Refined mansion tax proposal being fed into debate'
+skipping len=826 text='CHICAGO (STMW) — Three people were killed and at l'
+skipping len=980 text='SAN FRANCISCO – A new edition of an international '
+skipping len=311 text="Winter isn't done with us yet.\\n\\nOttawa can expect "
+skipping len=943 text='The Ice Light is “a portable, dimmable, daylight b'
+skipping len=188 text='A Wall Street sign is displayed in front of the Ne'
+[data] seen=10000 kept=3124 dropped=6876
+[data] seen=20000 kept=6318 dropped=13682
+[data] seen=30000 kept=9522 dropped=20478
+[data] seen=40000 kept=12678 dropped=27322
+[data] seen=50000 kept=15884 dropped=34116
+[data] seen=60000 kept=19035 dropped=40965
+[data] seen=70000 kept=22161 dropped=47839
+[data] seen=80000 kept=25353 dropped=54647
+[data] seen=90000 kept=28533 dropped=61467
+[data] seen=100000 kept=31656 dropped=68344
+[data] seen=110000 kept=34893 dropped=75107
+[data] seen=120000 kept=38138 dropped=81862
+[data] seen=130000 kept=41380 dropped=88620
+[data] seen=140000 kept=44554 dropped=95446
+[data] seen=150000 kept=47806 dropped=102194
+[data] seen=160000 kept=51016 dropped=108984
+[data] seen=170000 kept=54156 dropped=115844
+[data] seen=180000 kept=57380 dropped=122620
+[data] seen=190000 kept=60576 dropped=129424
+[data] seen=200000 kept=63722 dropped=136278
+[data] seen=210000 kept=66866 dropped=143134
+[data] seen=220000 kept=70014 dropped=149986
+[data] seen=230000 kept=73202 dropped=156798
+[data] seen=240000 kept=76353 dropped=163647

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/activate

@@ -0,0 +1,130 @@
+# Copyright (c) 2020-202x The virtualenv developers
+#
+# Permission is hereby granted, free of charge, to any person obtaining
+# a copy of this software and associated documentation files (the
+# "Software"), to deal in the Software without restriction, including
+# without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and to
+# permit persons to whom the Software is furnished to do so, subject to
+# the following conditions:
+#
+# The above copyright notice and this permission notice shall be
+# included in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+# EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+# LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+# OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+# WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# This file must be used with "source bin/activate" *from bash*
+# you cannot run it directly
+
+if ! [ -z "${SCRIPT_PATH+_}" ] ; then
+    _OLD_SCRIPT_PATH="$SCRIPT_PATH"
+fi
+
+# Get script path (only used if environment is relocatable).
+if [ -n "${BASH_VERSION:+x}" ] ; then
+    SCRIPT_PATH="${BASH_SOURCE[0]}"
+    if [ "$SCRIPT_PATH" = "$0" ]; then
+        # Only bash has a reasonably robust check for source'dness.
+        echo "You must source this script: \$ source $0" >&2
+        exit 33
+    fi
+elif [ -n "${ZSH_VERSION:+x}" ] ; then
+    SCRIPT_PATH="${(%):-%x}"
+elif [ -n "${KSH_VERSION:+x}" ] ; then
+    SCRIPT_PATH="${.sh.file}"
+fi
+
+deactivate () {
+    unset -f pydoc >/dev/null 2>&1 || true
+
+    # reset old environment variables
+    # ! [ -z ${VAR+_} ] returns true if VAR is declared at all
+    if ! [ -z "${_OLD_VIRTUAL_PATH:+_}" ] ; then
+        PATH="$_OLD_VIRTUAL_PATH"
+        export PATH
+        unset _OLD_VIRTUAL_PATH
+    fi
+    if ! [ -z "${_OLD_VIRTUAL_PYTHONHOME+_}" ] ; then
+        PYTHONHOME="$_OLD_VIRTUAL_PYTHONHOME"
+        export PYTHONHOME
+        unset _OLD_VIRTUAL_PYTHONHOME
+    fi
+
+    # The hash command must be called to get it to forget past
+    # commands. Without forgetting past commands the $PATH changes
+    # we made may not be respected
+    hash -r 2>/dev/null
+
+    if ! [ -z "${_OLD_VIRTUAL_PS1+_}" ] ; then
+        PS1="$_OLD_VIRTUAL_PS1"
+        export PS1
+        unset _OLD_VIRTUAL_PS1
+    fi
+
+    unset VIRTUAL_ENV
+    unset VIRTUAL_ENV_PROMPT
+    if [ ! "${1-}" = "nondestructive" ] ; then
+    # Self destruct!
+        unset -f deactivate
+    fi
+}
+
+# unset irrelevant variables
+deactivate nondestructive
+
+VIRTUAL_ENV='/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv'
+if ([ "$OSTYPE" = "cygwin" ] || [ "$OSTYPE" = "msys" ]) && $(command -v cygpath &> /dev/null) ; then
+    VIRTUAL_ENV=$(cygpath -u "$VIRTUAL_ENV")
+fi
+export VIRTUAL_ENV
+
+# Unset the `SCRIPT_PATH` variable, now that the `VIRTUAL_ENV` variable
+# has been set. This is important for relocatable environments.
+if ! [ -z "${_OLD_SCRIPT_PATH+_}" ] ; then
+    SCRIPT_PATH="$_OLD_SCRIPT_PATH"
+    export SCRIPT_PATH
+    unset _OLD_SCRIPT_PATH
+else
+    unset SCRIPT_PATH
+fi
+
+_OLD_VIRTUAL_PATH="$PATH"
+PATH="$VIRTUAL_ENV/bin:$PATH"
+export PATH
+
+if [ "x" != x ] ; then
+    VIRTUAL_ENV_PROMPT=""
+else
+    VIRTUAL_ENV_PROMPT=$(basename "$VIRTUAL_ENV")
+fi
+export VIRTUAL_ENV_PROMPT
+
+# unset PYTHONHOME if set
+if ! [ -z "${PYTHONHOME+_}" ] ; then
+    _OLD_VIRTUAL_PYTHONHOME="$PYTHONHOME"
+    unset PYTHONHOME
+fi
+
+if [ -z "${VIRTUAL_ENV_DISABLE_PROMPT-}" ] ; then
+    _OLD_VIRTUAL_PS1="${PS1-}"
+    PS1="(${VIRTUAL_ENV_PROMPT}) ${PS1-}"
+    export PS1
+fi
+
+# Make sure to unalias pydoc if it's already there
+alias pydoc 2>/dev/null >/dev/null && unalias pydoc || true
+
+pydoc () {
+    python -m pydoc "$@"
+}
+
+# The hash command must be called to get it to forget past
+# commands. Without forgetting past commands the $PATH changes
+# we made may not be respected
+hash -r 2>/dev/null || true

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/activate.ps1

@@ -0,0 +1,82 @@
+# Copyright (c) 2020-202x The virtualenv developers
+#
+# Permission is hereby granted, free of charge, to any person obtaining
+# a copy of this software and associated documentation files (the
+# "Software"), to deal in the Software without restriction, including
+# without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and to
+# permit persons to whom the Software is furnished to do so, subject to
+# the following conditions:
+#
+# The above copyright notice and this permission notice shall be
+# included in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+# EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+# LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+# OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+# WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+$script:THIS_PATH = $myinvocation.mycommand.path
+$script:BASE_DIR = Split-Path (Resolve-Path "$THIS_PATH/..") -Parent
+
+function global:deactivate([switch] $NonDestructive) {
+    if (Test-Path variable:_OLD_VIRTUAL_PATH) {
+        $env:PATH = $variable:_OLD_VIRTUAL_PATH
+        Remove-Variable "_OLD_VIRTUAL_PATH" -Scope global
+    }
+
+    if (Test-Path function:_old_virtual_prompt) {
+        $function:prompt = $function:_old_virtual_prompt
+        Remove-Item function:\_old_virtual_prompt
+    }
+
+    if ($env:VIRTUAL_ENV) {
+        Remove-Item env:VIRTUAL_ENV -ErrorAction SilentlyContinue
+    }
+
+    if ($env:VIRTUAL_ENV_PROMPT) {
+        Remove-Item env:VIRTUAL_ENV_PROMPT -ErrorAction SilentlyContinue
+    }
+
+    if (!$NonDestructive) {
+        # Self destruct!
+        Remove-Item function:deactivate
+        Remove-Item function:pydoc
+    }
+}
+
+function global:pydoc {
+    python -m pydoc $args
+}
+
+# unset irrelevant variables
+deactivate -nondestructive
+
+$VIRTUAL_ENV = $BASE_DIR
+$env:VIRTUAL_ENV = $VIRTUAL_ENV
+
+if ("" -ne "") {
+    $env:VIRTUAL_ENV_PROMPT = ""
+}
+else {
+    $env:VIRTUAL_ENV_PROMPT = $( Split-Path $env:VIRTUAL_ENV -Leaf )
+}
+
+New-Variable -Scope global -Name _OLD_VIRTUAL_PATH -Value $env:PATH
+
+$env:PATH = "$env:VIRTUAL_ENV/bin:" + $env:PATH
+if (!$env:VIRTUAL_ENV_DISABLE_PROMPT) {
+    function global:_old_virtual_prompt {
+        ""
+    }
+    $function:_old_virtual_prompt = $function:prompt
+
+    function global:prompt {
+        # Add the custom prefix to the existing prompt
+        $previous_prompt_value = & $function:_old_virtual_prompt
+        ("(" + $env:VIRTUAL_ENV_PROMPT + ") " + $previous_prompt_value)
+    }
+}

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/activate_this.py

@@ -0,0 +1,59 @@
+# Copyright (c) 2020-202x The virtualenv developers
+#
+# Permission is hereby granted, free of charge, to any person obtaining
+# a copy of this software and associated documentation files (the
+# "Software"), to deal in the Software without restriction, including
+# without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and to
+# permit persons to whom the Software is furnished to do so, subject to
+# the following conditions:
+#
+# The above copyright notice and this permission notice shall be
+# included in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+# EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+# LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+# OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+# WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+"""
+Activate virtualenv for current interpreter:
+
+import runpy
+runpy.run_path(this_file)
+
+This can be used when you must use an existing Python interpreter, not the virtualenv bin/python.
+"""  # noqa: D415
+
+from __future__ import annotations
+
+import os
+import site
+import sys
+
+try:
+    abs_file = os.path.abspath(__file__)
+except NameError as exc:
+    msg = "You must use import runpy; runpy.run_path(this_file)"
+    raise AssertionError(msg) from exc
+
+bin_dir = os.path.dirname(abs_file)
+base = bin_dir[: -len("bin") - 1]  # strip away the bin part from the __file__, plus the path separator
+
+# prepend bin to PATH (this file is inside the bin directory)
+os.environ["PATH"] = os.pathsep.join([bin_dir, *os.environ.get("PATH", "").split(os.pathsep)])
+os.environ["VIRTUAL_ENV"] = base  # virtual env is right above bin directory
+os.environ["VIRTUAL_ENV_PROMPT"] = "" or os.path.basename(base)  # noqa: SIM222
+
+# add the virtual environments libraries to the host python import mechanism
+prev_length = len(sys.path)
+for lib in "../lib/python3.12/site-packages".split(os.pathsep):
+    path = os.path.realpath(os.path.join(bin_dir, lib))
+    site.addsitedir(path)
+sys.path[:] = sys.path[prev_length:] + sys.path[0:prev_length]
+
+sys.real_prefix = sys.prefix
+sys.prefix = base

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/httpx

@@ -0,0 +1,10 @@
+#!/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/python3
+# -*- coding: utf-8 -*-
+import sys
+from httpx import main
+if __name__ == "__main__":
+    if sys.argv[0].endswith("-script.pyw"):
+        sys.argv[0] = sys.argv[0][:-11]
+    elif sys.argv[0].endswith(".exe"):
+        sys.argv[0] = sys.argv[0][:-4]
+    sys.exit(main())

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/python

@@ -0,0 +1 @@
+/usr/bin/python

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/python3.12

@@ -0,0 +1 @@
+python

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/tiny-agents

@@ -0,0 +1,10 @@
+#!/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/python3
+# -*- coding: utf-8 -*-
+import sys
+from huggingface_hub.inference._mcp.cli import app
+if __name__ == "__main__":
+    if sys.argv[0].endswith("-script.pyw"):
+        sys.argv[0] = sys.argv[0][:-11]
+    elif sys.argv[0].endswith(".exe"):
+        sys.argv[0] = sys.argv[0][:-4]
+    sys.exit(app())

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/tqdm

@@ -0,0 +1,10 @@
+#!/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/bin/python3
+# -*- coding: utf-8 -*-
+import sys
+from tqdm.cli import main
+if __name__ == "__main__":
+    if sys.argv[0].endswith("-script.pyw"):
+        sys.argv[0] = sys.argv[0][:-11]
+    elif sys.argv[0].endswith(".exe"):
+        sys.argv[0] = sys.argv[0][:-4]
+    sys.exit(main())

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/evolla/modular_evolla.py

@@ -0,0 +1,893 @@
+# Copyright 2025 Westlake Representational Learning Lab (Fajie Yuan Lab) team 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.
+
+from dataclasses import dataclass
+
+import torch
+from torch import nn
+
+from ... import initialization as init
+from ...cache_utils import Cache, DynamicCache
+from ...generation import GenerationMixin
+from ...masking_utils import create_bidirectional_mask, create_causal_mask
+from ...modeling_outputs import (
+    BaseModelOutputWithPast,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithPast,
+    ModelOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    auto_docstring,
+    can_return_tuple,
+    logging,
+)
+from ...utils.generic import merge_with_config_defaults
+from ...utils.output_capturing import OutputRecorder, capture_outputs
+from ..esm.modeling_esm import (
+    EsmAttention,
+    EsmEmbeddings,
+    EsmEncoder,
+    EsmIntermediate,
+    EsmLayer,
+    EsmOutput,
+    EsmPooler,
+    EsmRotaryEmbedding,
+    EsmSelfAttention,
+    EsmSelfOutput,
+)
+from ..llama.modeling_llama import (
+    LlamaAttention,
+    LlamaDecoderLayer,
+    LlamaMLP,
+    LlamaPreTrainedModel,
+    LlamaRMSNorm,
+    LlamaRotaryEmbedding,
+)
+from .configuration_evolla import EvollaConfig, SaProtConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+class EvollaSaProtEmbeddings(EsmEmbeddings):
+    def __init__(self, config):
+        super().__init__(config)
+        # remove the position_ids in EsmEmbeddings
+        self.position_ids = None
+
+
+class EvollaSaProtRotaryEmbedding(EsmRotaryEmbedding):
+    def __init__(self, config: SaProtConfig, device=None):
+        super().__init__(config, device)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: SaProtConfig | None = None,
+        device: "torch.device | None" = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        return super().compute_default_rope_parameters(config, device, seq_len)
+
+
+class EvollaSaProtSelfAttention(EsmSelfAttention):
+    pass
+
+
+class EvollaSaProtSelfOutput(EsmSelfOutput):
+    pass
+
+
+class EvollaSaProtAttention(EsmAttention):
+    pass
+
+
+class EvollaSaProtIntermediate(EsmIntermediate):
+    pass
+
+
+class EvollaSaProtOutput(EsmOutput):
+    pass
+
+
+class EvollaSaProtLayer(EsmLayer):
+    pass
+
+
+class EvollaSaProtEncoder(EsmEncoder):
+    pass
+
+
+class EvollaSaProtPooler(EsmPooler):
+    pass
+
+
+@auto_docstring
+class EvollaSaProtPreTrainedModel(PreTrainedModel):
+    config: SaProtConfig
+    _no_split_modules = ["EvollaSaProtLayer"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_attention_backend = True
+
+    _can_record_outputs = {
+        "hidden_states": EvollaSaProtLayer,
+        "attentions": [OutputRecorder(EvollaSaProtSelfAttention, index=1, layer_name="attention")],
+        "cross_attentions": [
+            OutputRecorder(EvollaSaProtSelfAttention, index=1, layer_name="crossattention"),
+        ],
+    }
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, EvollaSaProtRotaryEmbedding):
+            curr_inv_freq, _ = module.compute_default_rope_parameters(module.config)
+            init.copy_(getattr(module, "inv_freq"), curr_inv_freq)
+
+
+class EvollaSaProtProteinEncoder(EvollaSaProtPreTrainedModel):
+    def __init__(self, config: SaProtConfig):
+        super().__init__(config)
+        self.embeddings = EvollaSaProtEmbeddings(config)
+        self.rotary_embeddings = EvollaSaProtRotaryEmbedding(config=config)
+        self.encoder = EvollaSaProtEncoder(config)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    @merge_with_config_defaults
+    @capture_outputs
+    def forward(
+        self,
+        input_ids: torch.Tensor | None,
+        attention_mask: torch.Tensor | None = None,
+        **kwargs,
+    ) -> tuple[torch.Tensor] | BaseModelOutputWithPoolingAndCrossAttentions:
+        input_shape = input_ids.size()
+        batch_size, seq_length = input_shape
+
+        device = input_ids.device
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length)), device=device)
+        inputs_embeds = self.embeddings(input_ids=input_ids, attention_mask=attention_mask)
+
+        attention_mask = create_bidirectional_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+        )
+
+        position_ids = torch.arange(seq_length, device=device).unsqueeze(0)
+        position_embeddings = self.rotary_embeddings(inputs_embeds, position_ids)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds, attention_mask=attention_mask, position_embeddings=position_embeddings, **kwargs
+        )
+        sequence_output = encoder_outputs[0]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class EvollaSequenceCompressorAttention(nn.Module):
+    def __init__(self, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm_media = nn.LayerNorm(dim)
+        self.norm_latents = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents, mask):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D);  n2: num of latent tokens
+        """
+        x = self.norm_media(x)
+        latents = self.norm_latents(latents)
+
+        h = self.heads
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(
+            2, dim=-1
+        )  # each: batch_size, max_protein_length+num_latents, dim_head*num_heads
+
+        q = q.view(q.size(0), q.size(1), h, -1).permute(0, 2, 1, 3)
+        k = k.view(k.size(0), k.size(1), h, -1).permute(0, 2, 1, 3)
+        v = v.view(v.size(0), v.size(1), h, -1).permute(0, 2, 1, 3)
+        q = q * self.scale  # batch_size, num_heads, num_latents, dim_head
+
+        # attention
+        sim = torch.matmul(q, k.transpose(-1, -2))
+        sim = sim - sim.amax(dim=-1, keepdim=True).detach()
+        bs, nh, skd, okd = sim.shape
+        ones = torch.ones(nh, skd).to(mask.device)  # Create a tensor of ones with shape (nh, skd)
+        mask_exp = mask[:, None, None, :]
+        ones_exp = ones[None, :, :, None]
+        mask = mask_exp * ones_exp
+
+        sim = sim.masked_fill((1 - mask).bool(), -1e4)
+        attn = sim.softmax(dim=-1)
+        out = torch.matmul(attn, v)
+        out = out.permute(0, 2, 1, 3)
+
+        # [batch, seq, head, features] -> [batch, seq, head*features]
+        out = out.reshape(out.size(0), out.size(1), -1)
+
+        return self.to_out(out)
+
+
+class EvollaFeedForward(nn.Module):
+    def __init__(self, dim, mult=4):
+        super().__init__()
+        inner_dim = int(dim * mult)
+
+        self.norm = nn.LayerNorm(dim)
+        self.fc1 = nn.Linear(dim, inner_dim, bias=False)
+        self.activation = nn.GELU()
+        self.fc2 = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x):
+        return self.fc2(self.activation(self.fc1(self.norm(x))))
+
+
+class EvollaSequenceCompressorResampler(nn.Module):
+    def __init__(self, config: EvollaConfig):
+        super().__init__()
+        protein_repr_dim = config.protein_encoder_config.hidden_size
+        self.num_latents = config.resampler_num_latents
+        self.latents = nn.Parameter(torch.randn(self.num_latents, protein_repr_dim), requires_grad=True)
+        self.layers = nn.ModuleList([])
+        for _ in range(config.resampler_depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        EvollaSequenceCompressorAttention(
+                            dim=protein_repr_dim, dim_head=config.resampler_dim_head, heads=config.resampler_heads
+                        ),
+                        EvollaFeedForward(dim=protein_repr_dim, mult=config.resampler_ff_mult),
+                    ]
+                )
+            )
+
+        self.norm = nn.LayerNorm(config.hidden_size)
+        self.protein_projector = nn.Linear(protein_repr_dim, config.hidden_size)
+
+    def forward(self, embeds, mask):
+        b = embeds.shape[0]
+
+        bs, _ = mask.shape  # bs, max_protein_length
+        latent_mask = torch.ones(bs, self.num_latents).to(mask.device)
+        mask = torch.cat((mask, latent_mask), dim=1)  # bs, max_protein_length + num_latents
+
+        # blocks
+        ones = torch.ones(b).to(self.latents.device)
+        latents = self.latents[None] * ones.view(-1, 1, 1)  # [b,n,d]
+        latents = latents.to(embeds.dtype)
+        for attn, ff in self.layers:
+            latents = attn(embeds, latents, mask) + latents
+            latents = ff(latents) + latents
+
+        transformed_feature = self.protein_projector(latents)
+
+        return self.norm(transformed_feature)
+
+
+@auto_docstring
+@dataclass
+class EvollaProteinEncoderModelOutput(ModelOutput):
+    r"""
+    sequence_compressor_output (`torch.FloatTensor` of shape `(batch_size, compressed_seq_len, hidden_size)`, *optional*):
+        Compressed sequence representation produced by the sequence compressor module. The sequence length is
+        reduced from the original input length to `compressed_seq_len` via learned compression.
+    """
+
+    sequence_compressor_output: torch.FloatTensor | None = None
+    last_hidden_state: torch.FloatTensor | None = None
+    hidden_states: tuple[torch.FloatTensor, ...] | None = None
+    attentions: tuple[torch.FloatTensor, ...] | None = None
+
+
+class EvollaProteinEncoder(nn.Module):
+    def __init__(self, config: EvollaConfig):
+        super().__init__()
+        self.model = EvollaSaProtProteinEncoder(config=config.protein_encoder_config)
+        self.sequence_compressor_resampler = EvollaSequenceCompressorResampler(config=config)
+
+    @can_return_tuple
+    def forward(self, input_ids: torch.LongTensor, attention_mask: torch.FloatTensor, **kwargs):
+        protein_output = self.model(input_ids=input_ids, attention_mask=attention_mask)
+        protein_embeds = protein_output.last_hidden_state
+        sequence_repr = self.sequence_compressor_resampler(protein_embeds, attention_mask)
+
+        return EvollaProteinEncoderModelOutput(
+            sequence_compressor_output=sequence_repr,
+            last_hidden_state=protein_output.last_hidden_state,
+        )
+
+
+class EvollaSequenceAlignerCrossAttention(nn.Module):
+    def __init__(
+        self,
+        config,
+        protein_encoder_dim: int | None = None,
+        structure_encoder_dim: int | None = None,
+        msa_encoder_dim: int | None = None,
+    ):
+        super().__init__()
+
+        self.hidden_size = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.scale = self.num_attention_heads**-0.5
+        self.attention_head_size = int(self.hidden_size / self.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        attention_probs_dropout_prob = config.aligner_attention_probs_dropout_prob
+        enable_bias = config.aligner_enable_bias
+        ffn_mult = config.aligner_ffn_mult
+
+        self.query = nn.Linear(self.hidden_size, self.all_head_size)
+        if protein_encoder_dim is not None:
+            self.key_protein = nn.Linear(protein_encoder_dim, self.all_head_size)
+            self.value_protein = nn.Linear(protein_encoder_dim, self.all_head_size)
+        else:
+            self.key_protein = None
+            self.value_protein = None
+
+        if structure_encoder_dim is not None:
+            self.key_structure = nn.Linear(structure_encoder_dim, self.all_head_size)
+            self.value_structure = nn.Linear(structure_encoder_dim, self.all_head_size)
+        else:
+            self.key_structure = None
+            self.value_structure = None
+
+        if msa_encoder_dim is not None:
+            self.key_msa = nn.Linear(msa_encoder_dim, self.all_head_size)
+            self.value_msa = nn.Linear(msa_encoder_dim, self.all_head_size)
+        else:
+            self.key_msa = None
+            self.value_msa = None
+
+        self.attention_norm = EvollaRMSNorm(self.hidden_size)
+
+        self.dropout = nn.Dropout(attention_probs_dropout_prob)
+
+        self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=enable_bias)
+
+        self.ff = EvollaFeedForward(self.hidden_size, ffn_mult)
+        self.gate_attention = nn.Parameter(torch.tensor([0.0]))
+        self.gate_ffw = nn.Parameter(torch.tensor([0.0]))
+
+    def cross_attention(
+        self,
+        query_states,
+        protein_key_value_states,
+        structure_key_value_states,
+        msa_key_value_states,
+        query_attn_mask,
+        protein_kv_attn_mask,
+        structure_kv_attn_mask,
+        msa_kv_attn_mask,
+    ):
+        """
+        query_states: text
+        key_value_states: protein
+        query_states: [bs, query_seq_len, dim]
+        key_value_states: [bs, kv_seq_len, dim]
+        query_attn_mask: [bs, query_seq_len]
+        kv_attn_mask: [bs, kv_seq_len]
+        """
+
+        # Concatenate protein and structure
+        kv_attn_mask = [protein_kv_attn_mask, structure_kv_attn_mask, msa_kv_attn_mask]
+        kv_attn_mask = [_ for _ in kv_attn_mask if _ is not None]
+        if not kv_attn_mask:
+            raise ValueError("At least one modality should be provided for cross attention.")
+        kv_attn_mask = torch.cat(kv_attn_mask, dim=1)
+
+        query_layer = self.attention_norm(query_states)
+
+        # Warning: This place might cause issues, refers to
+        # https://discuss.pytorch.org/t/cuda-error-cublas-status-not-supported-when-calling-cublasltmatmul-from-torch-nn-functional-linear/170214/13
+        # Solution: add `DISABLE_ADDMM_CUDA_LT=1` as environment variable
+        # Apply linear transformation to input_query, input_key, and input_value
+        query_layer = self.query(query_layer)  # [bs, querylength, dim]
+
+        if self.key_protein is not None and self.value_protein is not None:
+            protein_key_value_states = protein_key_value_states.to(query_states)
+            key_layer_protein = self.key_protein(protein_key_value_states)  # [bs, keylength, dim]
+            value_layer_protein = self.value_protein(protein_key_value_states)  # [bs, keylength, dim]
+        else:
+            key_layer_protein = None
+            value_layer_protein = None
+
+        if self.key_structure is not None and self.value_structure is not None:
+            structure_key_value_states = structure_key_value_states.to(query_states)
+            key_layer_structure = self.key_structure(structure_key_value_states)  # [bs, keylength, dim]
+            value_layer_structure = self.value_structure(structure_key_value_states)  # [bs, keylength, dim]
+        else:
+            key_layer_structure = None
+            value_layer_structure = None
+
+        if self.key_msa is not None and self.value_msa is not None:
+            msa_key_value_states = msa_key_value_states.to(query_states)
+            key_layer_msa = self.key_msa(msa_key_value_states)  # [bs, keylength, dim]
+            value_layer_msa = self.value_msa(msa_key_value_states)  # [bs, keylength, dim]
+        else:
+            key_layer_msa = None
+            value_layer_msa = None
+
+        key_layer = [key_layer_protein, key_layer_structure, key_layer_msa]
+        key_layer = [_ for _ in key_layer if _ is not None]
+        key_layer = torch.cat(key_layer, dim=1)
+
+        value_layer = [value_layer_protein, value_layer_structure, value_layer_msa]
+        value_layer = [_ for _ in value_layer if _ is not None]
+        value_layer = torch.cat(value_layer, dim=1)
+
+        new_query_layer_shape = query_layer.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        query_layer = query_layer.view(*new_query_layer_shape).permute(0, 2, 1, 3)
+
+        new_key_layer_shape = key_layer.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        key_layer = key_layer.view(*new_key_layer_shape).permute(0, 2, 1, 3)
+
+        new_value_layer_shape = value_layer.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        value_layer = value_layer.view(*new_value_layer_shape).permute(0, 2, 1, 3)
+
+        query_layer = query_layer * self.scale
+
+        # attention_mask: [bs, 1, querylength, keylength]
+        if query_attn_mask is None:
+            query_attn_mask = torch.ones(query_states.size(0), query_states.size(1)).to(query_states.device)
+        attention_mask = query_attn_mask[:, None, :, None] * kv_attn_mask[:, None, None, :]
+        # Compute the scaled dot-product attention scores
+        attn_weights = torch.matmul(query_layer, key_layer.transpose(-1, -2))  # [bs, numheads, querylength, keylength]
+        attn_weights = attn_weights - attn_weights.amax(dim=-1, keepdim=True).detach()  # To stabilize score
+        attention_scores = attn_weights.masked_fill(
+            (1 - attention_mask).bool(), torch.finfo(attn_weights.dtype).min
+        )  # [bs, numheads, querylength, keylength]
+
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # attention_probs_dropped = self.dropout(attention_probs)
+
+        context_layer = torch.matmul(attention_probs, value_layer)  # [bs, numheads, querylength, dim/numheads]
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        context_layer = self.out_proj(context_layer)
+
+        return context_layer
+
+    def forward(
+        self,
+        query_states,
+        protein_kv_states,
+        structure_kv_states,
+        msa_kv_states,
+        query_attn_mask,
+        protein_kv_attn_mask=None,
+        structure_kv_attn_mask=None,
+        msa_kv_attn_mask=None,
+        protein_batch_mask=None,
+        structure_batch_mask=None,
+        msa_batch_mask=None,
+        past_key_values=None,
+    ):
+        if protein_kv_states is not None:
+            bs, protein_kv_seq_len, dim = protein_kv_states.shape
+            if protein_kv_attn_mask is None:
+                protein_kv_attn_mask = (
+                    torch.ones(bs, protein_kv_seq_len).to(protein_batch_mask.device)
+                    * protein_batch_mask.expand(size=(protein_kv_seq_len, bs)).T
+                ).to(protein_kv_states.device)
+        else:
+            protein_kv_attn_mask = None
+
+        if structure_kv_states is not None:
+            bs, structure_kv_seq_len, dim = structure_kv_states.shape
+            if structure_kv_attn_mask is None:
+                structure_kv_attn_mask = (
+                    torch.ones(bs, structure_kv_seq_len).to(protein_batch_mask.device)
+                    * structure_batch_mask.expand(size=(structure_kv_seq_len, bs)).T
+                ).to(structure_kv_states.device)
+        else:
+            structure_kv_attn_mask = None
+
+        if msa_kv_states is not None:
+            bs, msa_kv_seq_len, dim = msa_kv_states.shape
+            if msa_kv_attn_mask is None:
+                msa_kv_attn_mask = (
+                    torch.ones(bs, msa_kv_seq_len).to(protein_batch_mask.device)
+                    * msa_batch_mask.expand(size=(msa_kv_seq_len, bs)).T
+                ).to(msa_kv_states.device)
+        else:
+            msa_kv_attn_mask = None
+        hidden_states = query_states
+        # only when there's at least one valid modality, crossattention will be performed
+        if (
+            (protein_kv_states is not None and protein_kv_attn_mask.any())
+            or (structure_kv_states is not None and structure_kv_attn_mask.any())
+            or (msa_kv_states is not None and msa_kv_attn_mask.any())
+        ):
+            residual = hidden_states
+            hidden_states = self.cross_attention(
+                query_states=hidden_states,
+                protein_key_value_states=protein_kv_states,
+                structure_key_value_states=structure_kv_states,
+                msa_key_value_states=msa_kv_states,
+                query_attn_mask=query_attn_mask,
+                protein_kv_attn_mask=protein_kv_attn_mask,
+                structure_kv_attn_mask=structure_kv_attn_mask,
+                msa_kv_attn_mask=msa_kv_attn_mask,
+            )  # [bs, query_seq_len, dim]
+            # tanh gate
+            hidden_states = torch.tanh(self.gate_attention) * hidden_states
+
+            hidden_states = residual + hidden_states  # input_query
+
+            residual = hidden_states
+            hidden_states = self.ff(hidden_states) * torch.tanh(self.gate_ffw)
+            hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class EvollaRMSNorm(LlamaRMSNorm):
+    pass
+
+
+class EvollaRotaryEmbedding(LlamaRotaryEmbedding):
+    pass
+
+
+class EvollaMLP(LlamaMLP):
+    pass
+
+
+class EvollaAttention(LlamaAttention):
+    pass
+
+
+class EvollaDecoderLayer(LlamaDecoderLayer):
+    def __init__(self, config: EvollaConfig, layer_idx: int):
+        super().__init__(config, layer_idx)
+        if (layer_idx + 1) % max(config.num_hidden_layers // config.aligner_num_add_layers, 1) == 0:
+            self.adapter = EvollaSequenceAlignerCrossAttention(
+                config,
+                protein_encoder_dim=config.hidden_size,
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        protein_kv_states: torch.Tensor | None = None,
+        structure_kv_states: torch.Tensor | None = None,
+        msa_kv_states: torch.Tensor | None = None,
+        protein_batch_mask: torch.Tensor | None = None,
+        structure_batch_mask: torch.Tensor | None = None,
+        msa_batch_mask: torch.Tensor | None = None,
+        query_attn_mask: torch.Tensor | None = None,
+        **kwargs,
+    ):
+        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
+
+        if hasattr(self, "adapter"):
+            hidden_states = self.adapter(
+                query_states=hidden_states,
+                protein_kv_states=protein_kv_states,
+                structure_kv_states=structure_kv_states,
+                msa_kv_states=msa_kv_states,
+                query_attn_mask=query_attn_mask,
+                protein_batch_mask=protein_batch_mask,
+                structure_batch_mask=structure_batch_mask,
+                msa_batch_mask=msa_batch_mask,
+            )
+
+        return hidden_states
+
+
+class EvollaPreTrainedModel(LlamaPreTrainedModel):
+    _supports_flash_attn = False  # see dependency on `EvollaSequenceCompressorResampler`
+    _supports_flex_attn = False  # see dependency on `EvollaSequenceCompressorResampler`
+    _supports_attention_backend = False
+    _no_split_modules = [
+        "EvollaDecoderLayer",
+        "EvollaSaProtLayer",
+        "EvollaSequenceCompressorResampler",
+        "EvollaSequenceAlignerCrossAttention",
+    ]
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        PreTrainedModel._init_weights(self, module)
+        if isinstance(module, EvollaSequenceAlignerCrossAttention):
+            init.zeros_(module.gate_attention)
+            init.zeros_(module.gate_ffw)
+            init.ones_(module.attention_norm.weight)
+        elif isinstance(module, EvollaSequenceCompressorResampler):
+            init.normal_(module.latents, mean=0.0, std=std)
+
+
+class EvollaModel(EvollaPreTrainedModel):
+    def __init__(self, config: EvollaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(self.vocab_size, config.hidden_size, self.padding_idx)
+        self.protein_encoder = EvollaProteinEncoder(config=config)
+        self.layers = nn.ModuleList(
+            [
+                EvollaDecoderLayer(
+                    config=config,
+                    layer_idx=layer_idx,
+                )
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+
+        self.norm = EvollaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = getattr(config, "gradient_checkpointing", False)
+        self.rotary_emb = EvollaRotaryEmbedding(config=config)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @auto_docstring
+    @merge_with_config_defaults
+    @capture_outputs
+    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,
+        protein_input_ids: torch.LongTensor | None = None,
+        protein_attention_mask: torch.Tensor | None = None,
+        structure_feats: torch.FloatTensor | None = None,
+        msa_feats: torch.FloatTensor | None = None,
+        structure_batch_mask: torch.Tensor | None = None,
+        msa_batch_mask: torch.Tensor | None = None,
+        **kwargs,
+    ) -> tuple | BaseModelOutputWithPast:
+        r"""
+        protein_input_ids (torch.LongTensor):
+            The input IDs for the protein sequence in structure-aware tokens. Should be of shape `(batch_size, protein_seq_length)` and type `torch.LongTensor`.
+        protein_attention_mask (torch.Tensor):
+            The attention mask for the protein sequence. Should be of shape `(batch_size, protein_seq_length)` and type `torch.Tensor`.
+        structure_feats (torch.FloatTensor):
+            The input IDs for purely structure-based features. Should be of shape `(batch_size, structure_seq_length, structure_feat_dim)` and type `torch.FloatTensor`. Dummy input for now.
+        msa_feats (torch.FloatTensor):
+            The input IDs for purely MSA-based features. Should be of shape `(batch_size, msa_seq_length, msa_feat_dim)` and type `torch.FloatTensor`. Dummy input for now.
+        structure_batch_mask (torch.Tensor):
+            The batch mask to decide which protein sequences are purely structure-based. Should be of shape `(batch_size)` and type `torch.Tensor`. Should be paired with `structure_feats`. Dummpy input for now.
+        msa_batch_mask (torch.Tensor):
+            The batch mask to decide which protein sequences are purely MSA-based. Should be of shape `(batch_size)` and type `torch.Tensor`. Should be paired with `msa_feats`. Dummpy input for now.
+        """
+        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)
+
+        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)
+
+        protein_feats = None
+        protein_batch_mask = None
+        # If provided, actually compute them
+        if protein_input_ids is not None and protein_attention_mask is not None:
+            protein_outputs = self.protein_encoder(
+                input_ids=protein_input_ids,
+                attention_mask=protein_attention_mask,
+            )
+            protein_feats = protein_outputs.sequence_compressor_output
+            protein_batch_mask = torch.ones(
+                protein_input_ids.shape[0],
+                device=protein_input_ids.device,
+                dtype=torch.bool,
+            )
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+        )
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+
+        for decoder_layer in self.layers:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                protein_kv_states=protein_feats,
+                structure_kv_states=structure_feats,
+                msa_kv_states=msa_feats,
+                protein_batch_mask=protein_batch_mask,
+                structure_batch_mask=structure_batch_mask,
+                msa_batch_mask=msa_batch_mask,
+                query_attn_mask=attention_mask,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+
+        output = BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+        return output
+
+
+class EvollaForProteinText2Text(EvollaPreTrainedModel, GenerationMixin):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = EvollaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, self.vocab_size, bias=False)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        return self.model.set_input_embeddings(value)
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,  # text input ids
+        attention_mask: torch.Tensor | None = None,  # text attention mask
+        inputs_embeds: torch.FloatTensor | None = None,  # text input embeddings
+        labels: torch.LongTensor | None = None,
+        protein_input_ids: torch.LongTensor | None = None,
+        protein_attention_mask: torch.Tensor | None = None,
+        use_cache: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs,
+    ):
+        r"""
+        protein_input_ids (torch.LongTensor):
+            The input IDs for the protein sequence. Should be of shape `(batch_size, protein_seq_length)` and type `torch.LongTensor`.
+        protein_attention_mask (torch.Tensor):
+            The attention mask for the protein sequence. Should be of shape `(batch_size, protein_seq_length)` and type `torch.Tensor`.
+
+        Example:
+
+        ```python
+        >>> from transformers import EvollaProcessor, EvollaForProteinText2Text
+        >>> model = EvollaForProteinText2Text.from_pretrained("westlake/Evolla-10B-hf")
+        >>> processor = EvollaProcessor.from_pretrained("westlake/Evolla-10B-hf")
+
+        >>> protein_information = {
+            "aa_seq": "your amino acid sequence",
+            "foldseek": "your foldseek sequence",
+        }
+        >>> question = "What is the function of this protein?"
+        >>> message = [
+            {"role": "system", "content": "You are an AI expert that can answer any questions about protein."},
+            {"role": "user", "content": question},
+        ]
+
+        >>> inputs = processor(proteins=[protein_information], messages_list=[message], return_tensors="pt", padding="longest")
+        >>> outputs = model.generate(**inputs)
+
+        >>> print(processor.batch_decode(outputs, skip_special_tokens=True))
+        ```"""
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            protein_input_ids=protein_input_ids,
+            protein_attention_mask=protein_attention_mask,
+            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.vocab_size, **kwargs)
+
+        lm_outputs = CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+        return lm_outputs
+
+
+__all__ = ["EvollaForProteinText2Text", "EvollaModel", "EvollaPreTrainedModel"]

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

@@ -0,0 +1,28 @@
+# Copyright 2025 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_vaultgemma import *
+    from .modeling_vaultgemma 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/vaultgemma/configuration_vaultgemma.py

@@ -0,0 +1,109 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/vaultgemma/modular_vaultgemma.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_vaultgemma.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2025 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 huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...modeling_rope_utils import RopeParameters
+from ...utils import auto_docstring
+
+
+@auto_docstring(checkpoint="google/vaultgemma-1b")
+@strict
+class VaultGemmaConfig(PreTrainedConfig):
+    r"""
+    query_pre_attn_scalar (`float`, *optional*, defaults to 256):
+        scaling factor used on the attention scores
+    final_logit_softcapping (`float`, *optional*, defaults to 30.0):
+        scaling factor when applying tanh softcapping on the logits.
+    attn_logit_softcapping (`float`, *optional*, defaults to 50.0):
+        scaling factor when applying tanh softcapping on the attention scores.
+
+    ```python
+    >>> from transformers import VaultGemmaModel, VaultGemmaConfig
+    >>> # Initializing a VaultGemma vaultgemma-7b style configuration
+    >>> configuration = VaultGemmaConfig()
+    >>> # Initializing a model from the vaultgemma-7b style configuration
+    >>> model = VaultGemmaModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "vaultgemma"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_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"]),
+    }
+
+    vocab_size: int = 256000
+    hidden_size: int = 2304
+    intermediate_size: int = 9216
+    num_hidden_layers: int = 26
+    num_attention_heads: int = 8
+    num_key_value_heads: int = 4
+    head_dim: int = 256
+    hidden_activation: str = "gelu_pytorch_tanh"
+    max_position_embeddings: int = 8192
+    initializer_range: float = 0.02
+    rms_norm_eps: float = 1e-6
+    use_cache: bool = True
+    pad_token_id: int | None = 0
+    eos_token_id: int | list[int] | None = 1
+    bos_token_id: int | None = 2
+    tie_word_embeddings: bool = True
+    rope_parameters: RopeParameters | dict | None = None
+    attention_bias: bool = False
+    attention_dropout: int | float | None = 0.0
+    query_pre_attn_scalar: int = 256
+    sliding_window: int | None = 4096
+    layer_types: list[str] | None = None
+    final_logit_softcapping: float | None = 30.0
+    attn_logit_softcapping: float | None = 50.0
+
+    def __post_init__(self, **kwargs):
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention" if bool((i + 1) % 2) else "full_attention" for i in range(self.num_hidden_layers)
+            ]
+
+        super().__post_init__(**kwargs)
+
+    def validate_architecture(self):
+        """Part of `@strict`-powered validation. Validates the architecture of the config."""
+        if self.hidden_size % self.num_attention_heads != 0:
+            raise ValueError(
+                f"The hidden size ({self.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({self.num_attention_heads})."
+            )
+
+
+__all__ = ["VaultGemmaConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/vaultgemma/modeling_vaultgemma.py

@@ -0,0 +1,546 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/vaultgemma/modular_vaultgemma.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_vaultgemma.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# Copyright 2025 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 collections.abc import Callable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from ... import initialization as init
+from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache
+from ...generation import GenerationMixin
+from ...integrations import use_kernel_func_from_hub, use_kernelized_func
+from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import 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 maybe_autocast, merge_with_config_defaults
+from ...utils.output_capturing import capture_outputs
+from .configuration_vaultgemma import VaultGemmaConfig
+
+
+class VaultGemmaRMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.zeros(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float())
+        # Llama does x.to(float16) * w whilst VaultGemma is (x * w).to(float16)
+        # See https://github.com/huggingface/transformers/pull/29402
+        output = output * (1.0 + self.weight.float())
+        return output.type_as(x)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+
+class VaultGemmaMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.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_activation]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+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,
+    dropout: float | int = 0.0,
+    scaling: float | None = None,
+    softcap: float | None = None,
+    **kwargs,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    if scaling is None:
+        scaling = module.head_dim**-0.5
+
+    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 softcap is not None:
+        attn_weights = attn_weights / softcap
+        attn_weights = torch.tanh(attn_weights)
+        attn_weights = attn_weights * softcap
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    # upcast attention to fp32
+    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
+
+
+@use_kernelized_func(apply_rotary_pos_emb)
+class VaultGemmaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: VaultGemmaConfig, layer_idx: int):
+        super().__init__()
+        self.layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = config.query_pre_attn_scalar**-0.5
+        self.attention_dropout = self.config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        self.attn_logit_softcapping = self.config.attn_logit_softcapping
+        self.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        attention_mask: torch.Tensor | None = None,
+        past_key_values: Cache | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_values is not None:
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+
+        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=self.attention_dropout if self.training else 0.0,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,
+            softcap=self.attn_logit_softcapping,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class VaultGemmaDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: VaultGemmaConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.config = config
+        self.self_attn = VaultGemmaAttention(config=config, layer_idx=layer_idx)
+        self.mlp = VaultGemmaMLP(config)
+        self.input_layernorm = VaultGemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.pre_feedforward_layernorm = VaultGemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        **kwargs,
+    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.pre_feedforward_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class VaultGemmaRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: VaultGemmaConfig, 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: VaultGemmaConfig | 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 VaultGemmaTextScaledWordEmbedding(nn.Embedding):
+    """
+    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float = 1.0):
+        super().__init__(num_embeddings, embedding_dim, padding_idx)
+        self.scalar_embed_scale = embed_scale
+        self.register_buffer("embed_scale", torch.tensor(embed_scale), persistent=False)
+
+    def forward(self, input_ids: torch.Tensor):
+        return super().forward(input_ids) * self.embed_scale.to(self.weight.dtype)
+
+
+@auto_docstring
+class VaultGemmaPreTrainedModel(PreTrainedModel):
+    config: VaultGemmaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["VaultGemmaDecoderLayer"]
+    _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": VaultGemmaDecoderLayer,
+        "attentions": VaultGemmaAttention,
+    }
+
+    @torch.no_grad()
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        # We initialize with 0s to be 1 centered as the RMSNorm here does (1 + weight)
+        if "RMSNorm" in module.__class__.__name__:
+            init.zeros_(module.weight)
+        elif isinstance(module, VaultGemmaTextScaledWordEmbedding):
+            init.constant_(module.embed_scale, module.scalar_embed_scale)
+
+
+@auto_docstring
+class VaultGemmaModel(VaultGemmaPreTrainedModel):
+    def __init__(self, config: VaultGemmaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        # VaultGemma3 downcasts the below to bfloat16, causing sqrt(3072)=55.4256 to become 55.5. See https://github.com/huggingface/transformers/pull/29402
+        self.embed_tokens = VaultGemmaTextScaledWordEmbedding(
+            config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=self.config.hidden_size**0.5
+        )
+        self.layers = nn.ModuleList(
+            [VaultGemmaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = VaultGemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = VaultGemmaRotaryEmbedding(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)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "inputs_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+                "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),
+            }
+
+        # embed positions
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask_mapping[self.config.layer_types[i]],
+                position_embeddings=position_embeddings,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class VaultGemmaForCausalLM(VaultGemmaPreTrainedModel, 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 = VaultGemmaModel(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, VaultGemmaForCausalLM
+
+        >>> model = VaultGemmaForCausalLM.from_pretrained("google/gemma-2-9b")
+        >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
+
+        >>> prompt = "What is your favorite condiment?"
+        >>> 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]
+        "What is your favorite condiment?"
+        ```"""
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        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, :])
+        if self.config.final_logit_softcapping is not None:
+            logits = logits / self.config.final_logit_softcapping
+            logits = torch.tanh(logits)
+            logits = logits * self.config.final_logit_softcapping
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = ["VaultGemmaForCausalLM", "VaultGemmaModel", "VaultGemmaPreTrainedModel"]

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

@@ -0,0 +1,29 @@
+# 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_vit import *
+    from .image_processing_pil_vit import *
+    from .image_processing_vit import *
+    from .modeling_vit 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/vit/configuration_vit.py

@@ -0,0 +1,72 @@
+# Copyright 2021 Google AI 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.
+"""ViT model configuration"""
+
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+
+
+@auto_docstring(checkpoint="google/vit-base-patch16-224")
+@strict
+class ViTConfig(PreTrainedConfig):
+    r"""
+    encoder_stride (`int`, *optional*, defaults to 16):
+        Factor to increase the spatial resolution by in the decoder head for masked image modeling.
+    pooler_output_size (`int`, *optional*):
+        Dimensionality of the pooler layer. If None, defaults to `hidden_size`.
+    pooler_act (`str`, *optional*, defaults to `"tanh"`):
+        The activation function to be used by the pooler.
+
+    Example:
+
+    ```python
+    >>> from transformers import ViTConfig, ViTModel
+
+    >>> # Initializing a ViT vit-base-patch16-224 style configuration
+    >>> configuration = ViTConfig()
+
+    >>> # Initializing a model (with random weights) from the vit-base-patch16-224 style configuration
+    >>> model = ViTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "vit"
+
+    hidden_size: int = 768
+    num_hidden_layers: int = 12
+    num_attention_heads: int = 12
+    intermediate_size: int = 3072
+    hidden_act: str = "gelu"
+    hidden_dropout_prob: float | int = 0.0
+    attention_probs_dropout_prob: float | int = 0.0
+    initializer_range: float = 0.02
+    layer_norm_eps: float = 1e-12
+    image_size: int | list[int] | tuple[int, int] = 224
+    patch_size: int | list[int] | tuple[int, int] = 16
+    num_channels: int = 3
+    qkv_bias: bool = True
+    encoder_stride: int = 16
+    pooler_output_size: int | None = None
+    pooler_act: str = "tanh"
+
+    def __post_init__(self, **kwargs):
+        self.pooler_output_size = self.pooler_output_size if self.pooler_output_size else self.hidden_size
+        super().__post_init__(**kwargs)
+
+
+__all__ = ["ViTConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/vit/image_processing_pil_vit.py

@@ -0,0 +1,30 @@
+# Copyright 2022 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.
+"""Image processor class for ViT."""
+
+from ...image_processing_backends import PilBackend
+from ...image_utils import IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD, PILImageResampling
+
+
+class ViTImageProcessorPil(PilBackend):
+    resample = PILImageResampling.BILINEAR
+    image_mean = IMAGENET_STANDARD_MEAN
+    image_std = IMAGENET_STANDARD_STD
+    size = {"height": 224, "width": 224}
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+
+
+__all__ = ["ViTImageProcessorPil"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/vit/image_processing_vit.py

@@ -0,0 +1,30 @@
+# Copyright 2022 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.
+"""Image processor class for ViT."""
+
+from ...image_processing_backends import TorchvisionBackend
+from ...image_utils import IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD, PILImageResampling
+
+
+class ViTImageProcessor(TorchvisionBackend):
+    resample = PILImageResampling.BILINEAR
+    image_mean = IMAGENET_STANDARD_MEAN
+    image_std = IMAGENET_STANDARD_STD
+    size = {"height": 224, "width": 224}
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+
+
+__all__ = ["ViTImageProcessor"]