Merge branch 'main_307c5238' into remove-attributes-from-processors-ydshieh

Author: ydshieh

.github/workflows/check_failed_tests.yml

@@ -41,9 +41,14 @@ env:
 
 jobs:
   check_new_failures:
-    name: " "
+    name: "Find commits for new failing tests"
+    strategy:
+      matrix:
+        run_idx: [1]
     runs-on:
       group: aws-g5-4xlarge-cache
+    outputs:
+      process: ${{ steps.check_file.outputs.process }}
     container:
       image: ${{ inputs.docker }}
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -54,14 +59,17 @@ jobs:
           path: /transformers/ci_results_${{ inputs.job }}
 
       - name: Check file
+        id: check_file
         working-directory: /transformers
         run: |
           if [ -f ci_results_${{ inputs.job }}/new_failures.json ]; then
             echo "`ci_results_${{ inputs.job }}/new_failures.json` exists, continue ..."
             echo "process=true" >> $GITHUB_ENV
+            echo "process=true" >> $GITHUB_OUTPUT
           else
             echo "`ci_results_${{ inputs.job }}/new_failures.json` doesn't exist, abort."
             echo "process=false" >> $GITHUB_ENV
+            echo "process=false" >> $GITHUB_OUTPUT
           fi
 
       - uses: actions/download-artifact@v4
@@ -118,6 +126,10 @@ jobs:
         run: |
           python3 utils/print_env.py
 
+      - name: Install pytest-flakefinder
+        if: ${{ env.process == 'true' }}
+        run: python3 -m pip install pytest-flakefinder
+
       - name: Show installed libraries and their versions
         working-directory: /transformers
         if: ${{ env.process == 'true' }}
@@ -126,25 +138,63 @@ jobs:
       - name: Check failed tests
         working-directory: /transformers
         if: ${{ env.process == 'true' }}
-        run: python3 utils/check_bad_commit.py --start_commit ${{ inputs.start_sha }} --end_commit ${{ env.END_SHA }} --file ci_results_${{ inputs.job }}/new_failures.json --output_file new_failures_with_bad_commit.json
+        run: python3 utils/check_bad_commit.py --start_commit ${{ inputs.start_sha }} --end_commit ${{ env.END_SHA }} --file ci_results_${{ inputs.job }}/new_failures.json --output_file new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
 
       - name: Show results
         working-directory: /transformers
         if: ${{ env.process == 'true' }}
         run: |
-          ls -l new_failures_with_bad_commit.json
-          cat new_failures_with_bad_commit.json
+          ls -l new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
+          cat new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}
+          path: /transformers/new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
+
+  process_new_failures_with_commit_info:
+    name: "process bad commit reports"
+    needs: check_new_failures
+    if: needs.check_new_failures.outputs.process == 'true'
+    runs-on:
+      group: aws-g5-4xlarge-cache
+    container:
+      image: ${{ inputs.docker }}
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: ci_results_${{ inputs.job }}
+          path: /transformers/ci_results_${{ inputs.job }}
 
-      - name: Checkout back
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: new_failures_with_bad_commit_${{ inputs.job }}*
+          path: /transformers/new_failures_with_bad_commit_${{ inputs.job }}
+          merge-multiple: true
+
+      - name: Check files
+        working-directory: /transformers
+        run: |
+          ls -la /transformers
+          ls -la /transformers/new_failures_with_bad_commit_${{ inputs.job }}
+
+      # Currently, we only run with a single runner by using `run_idx: [1]`. We might try to run with multiple runners
+      # to further reduce the false positive caused by flaky tests, which requires further processing to merge reports.
+      - name: Merge files
+        shell: bash
         working-directory: /transformers
-        if: ${{ env.process == 'true' }}
         run: |
-          git checkout ${{ inputs.start_sha }}
+          cp /transformers/new_failures_with_bad_commit_${{ inputs.job }}/new_failures_with_bad_commit_${{ inputs.job }}_1.json new_failures_with_bad_commit.json
+
+      - name: Update clone
+        working-directory: /transformers
+        run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
 
       - name: Process report
         shell: bash
         working-directory: /transformers
-        if: ${{ env.process == 'true' }}
         env:
           ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
           TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
@@ -156,7 +206,6 @@ jobs:
       - name: Process report
         shell: bash
         working-directory: /transformers
-        if: ${{ env.process == 'true' }}
         env:
           ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
           TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
@@ -171,13 +220,12 @@ jobs:
 
       - name: Prepare Slack report title
         working-directory: /transformers
-        if: ${{ env.process == 'true' }}
         run: |
           pip install slack_sdk
           echo "title=$(python3 -c 'import sys; sys.path.append("utils"); from utils.notification_service import job_to_test_map; ci_event = "${{ inputs.ci_event }}"; job = "${{ inputs.job }}"; test_name = job_to_test_map[job]; title = f"New failed tests of {ci_event}" + ":" + f" {test_name}"; print(title)')" >> $GITHUB_ENV
 
       - name: Send processed report
-        if: ${{ env.process == 'true' && !endsWith(env.REPORT_TEXT, '{}') }}
+        if: ${{ !endsWith(env.REPORT_TEXT, '{}') }}
         uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
         with:
           # Slack channel id, channel name, or user id to post message.

docker/transformers-all-latest-gpu/Dockerfile

@@ -24,7 +24,8 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
 # 1. Put several commands in a single `RUN` to avoid image/layer exporting issue. Could be revised in the future.
 # 2. Regarding `torch` part, We might need to specify proper versions for `torchvision` and `torchaudio`.
 #    Currently, let's not bother to specify their versions explicitly (so installed with their latest release versions).
-RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
+# 3. For `torchcodec<0.8`: this is quickly added as torch 2.9.0 + torchcodec 0.8.0 fails on our CI env. Need to remove later once they work.
+RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio "torchcodec<0.8" --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
 
 RUN python3 -m pip install --no-cache-dir -U timm
 

docs/source/en/perf_infer_gpu_multi.md

@@ -45,17 +45,21 @@ This guide shows how to enable tensor parallelism with Transformers and differen
 
 ## Partitioning a model
 
-Transformers supports tensor parallelism if a model has a `tp_plan`. Set `tp_plan="auto"` to automatically use a tensor parallelism plan based on a model's predefined configuration.
+Transformers supports tensor parallelism if a model has a `tp_plan`. There are two ways to partition a model.
+
+- Set `tp_plan="auto"` to automatically use a tensor parallelism plan based on a model's predefined configuration.
+- Define and pass a manual `tp_plan`.
+
+<hfoptions id="tp_plan">
+<hfoption id="auto plan">
 
 ```py
 import os
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 # model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct" # better to visualize all the possible strategies
-model_id = "meta-llama/Meta-Llama-3-8B-Instruct"  # better for smaller number of GPUs
-
-model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, tp_plan="auto")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct" , dtype=torch.bfloat16, tp_plan="auto")
 print(model._tp_plan)
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")
@@ -72,6 +76,31 @@ Launch the inference script above on [torchrun](https://pytorch.org/docs/stable/
 torchrun --nproc-per-node 4 demo.py
 ```
 
+</hfoption>
+<hfoption id="manual plan">
+
+Define a tensor parallel plan for each layer in `tp_plan` and pass it to [`~PreTrainedModel.from_pretrained`]. The example below uses column and row partitioning. See the [Partitioning strategies](#partitioning-strategies) section for other supported strategies.
+
+Manual partitioning requires deep understanding of model architecture and strategy interactions. Poor partitioning choices create slow models that fail or produce incorrect results. The [Ultra-Scale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=tensor_parallelism) explains partitioning strategies in detail.
+
+```py
+from transformers import AutoModelForCausalLM
+
+tp_plan = {
+    "model.layers.*.self_attn.q_proj": "colwise",
+    "model.layers.*.self_attn.k_proj": "colwise",
+    "model.layers.*.self_attn.v_proj": "colwise",
+    "model.layers.*.self_attn.o_proj": "rowwise",
+    ...
+}
+
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct", dtype="auto", tp_plan=tp_plan)
+print(model.tp_plan)
+```
+
+</hfoption>
+</hfoptions>
+
 ## Partitioning strategies
 
 All partitioning strategies are defined in the [`ParallelInterface`] class which maps a string to the strategy implementation. You don't need to interact with this class directly since all the strategies are set with `tp_plan` in [`~PreTrainedModel.from_pretrained`], but it is useful for checking what strategies are available.

docs/source/ja/perf_train_gpu_many.md

@@ -472,8 +472,6 @@ FlexFlowは、サンプル-オペレータ-属性-パラメータの4D並列化
 
 したがって、このフレームワークの約束は非常に魅力的です。選択したクラスタで30分間のシミュレーションを実行し、この特定の環境を最適に利用するための最良の戦略を提供します。部分を追加/削除/置換すると、それに対して実行して再最適化プランを作成します。その後、トレーニングできます。異なるセットアップには独自の最適化があります。
 
-🤗 Transformersの現在の状況: まだ統合されていません。すでに[transformers.utils.fx](https://github.com/huggingface/transformers/blob/master/src/transformers/utils/fx.py)を使用してモデルがFXトレース可能であるため、FlexFlowを動作させるために必要な手順を誰かが見つける必要があります。
-
 ## Which Strategy To Use When
 
 ここでは、どの並列化戦略をいつ使用するかの非常におおまかなアウトラインを示します。各リストの最初が通常よりも速いことが一般的です。

docs/source/ko/_toctree.yml

@@ -1057,7 +1057,7 @@
         title: FLAVA
       - local: model_doc/gemma3
         title: Gemma3
-      - local: in_translation
+      - local: model_doc/gemma3n
         title: Gemma3n
       - local: in_translation
         title: GIT