Bump min torch version to 2.3.0 (#1152)

* bump min torch * make random correct less likely * bf16 precisioin * fix * new dataset? * try again * debug * debug * pin hf hub * remove debug --------- Co-authored-by: Milo Cress <milo.cress@databricks.com>
mosaicml · May 2, 2024 · a3e0fb9 · a3e0fb9
1 parent 0d58f46
commit a3e0fb9
Show file tree

Hide file tree

Showing 19 changed files with 23 additions and 35 deletions.
diff --git a/.github/workflows/docker.yaml b/.github/workflows/docker.yaml
@@ -17,12 +17,6 @@ jobs:
     strategy:
       matrix:
         include:
-        - name: "2.2.1_cu121_flash2"
-          base_image: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04
-          dep_groups: "[gpu-flash2]"
-        - name: "2.2.1_cu121_flash2_aws"
-          base_image: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04-aws
-          dep_groups: "[gpu-flash2]"
         - name: "2.3.0_cu121_flash2"
           base_image: mosaicml/pytorch:2.3.0_cu121-python3.11-ubuntu20.04
           dep_groups: "[gpu-flash2]"

diff --git a/.github/workflows/pr-cpu.yaml b/.github/workflows/pr-cpu.yaml
@@ -19,10 +19,6 @@ jobs:
     strategy:
       matrix:
         include:
-        - name: "cpu-2.2.1"
-          container: mosaicml/pytorch:2.2.1_cpu-python3.11-ubuntu20.04
-          markers: "not gpu"
-          pytest_command: "coverage run -m pytest"
         - name: "cpu-2.3.0"
           container: mosaicml/pytorch:2.3.0_cpu-python3.11-ubuntu20.04
           markers: "not gpu"

diff --git a/.github/workflows/pr-gpu.yaml b/.github/workflows/pr-gpu.yaml
@@ -19,11 +19,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-        - name: "gpu-2.2.1"
-          container: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04
-          markers: "gpu"
-          pip_deps: "[all]"
-          pytest_command: "coverage run -m pytest"
         - name: "gpu-2.3.0"
           container: mosaicml/pytorch:2.3.0_cu121-python3.11-ubuntu20.04
           markers: "gpu"

diff --git a/README.md b/README.md
@@ -113,24 +113,24 @@ If you have success/failure using LLM Foundry on other systems, please let us kn
 
 | Device         | Torch Version | Cuda Version | Status                       |
 | -------------- | ------------- | ------------ | ---------------------------- |
-| A100-40GB/80GB | 2.2.1         | 12.1         | :white_check_mark: Supported |
-| H100-80GB      | 2.2.1         | 12.1         | :white_check_mark: Supported |
+| A100-40GB/80GB | 2.3.0         | 12.1         | :white_check_mark: Supported |
+| H100-80GB      | 2.3.0         | 12.1         | :white_check_mark: Supported |
 
 ## MosaicML Docker Images
 We highly recommend using our prebuilt Docker images. You can find them here: https://hub.docker.com/orgs/mosaicml/repositories.
 
 The `mosaicml/pytorch` images are pinned to specific PyTorch and CUDA versions, and are stable and rarely updated.
 
 The `mosaicml/llm-foundry` images are built with new tags upon every commit to the `main` branch.
-You can select a specific commit hash such as `mosaicml/llm-foundry:2.2.1_cu121_flash2-36ab1ba` or take the latest one using `mosaicml/llm-foundry:2.2.1_cu121_flash2-latest`.
+You can select a specific commit hash such as `mosaicml/llm-foundry:2.3.0_cu121_flash2-36ab1ba` or take the latest one using `mosaicml/llm-foundry:2.3.0_cu121_flash2-latest`.
 
 **Please Note:** The `mosaicml/llm-foundry` images do not come with the `llm-foundry` package preinstalled, just the dependencies. You will still need to `pip install llm-foundry` either from PyPi or from source.
 
 | Docker Image                                           | Torch Version | Cuda Version      | LLM Foundry dependencies installed? |
 | ------------------------------------------------------ | ------------- | ----------------- | ----------------------------------- |
-| `mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04`  | 2.2.1         | 12.1 (Infiniband) | No                                  |
-| `mosaicml/llm-foundry:2.2.1_cu121_flash2-latest`       | 2.2.1         | 12.1 (Infiniband) | Yes                                 |
-| `mosaicml/llm-foundry:2.2.1_cu121_flash2_aws-latest`   | 2.2.1         | 12.1 (EFA)        | Yes                                 |
+| `mosaicml/pytorch:2.3.0_cu121-python3.11-ubuntu20.04`  | 2.3.0         | 12.1 (Infiniband) | No                                  |
+| `mosaicml/llm-foundry:2.3.0_cu121_flash2-latest`       | 2.3.0         | 12.1 (Infiniband) | Yes                                 |
+| `mosaicml/llm-foundry:2.3.0_cu121_flash2_aws-latest`   | 2.3.0         | 12.1 (EFA)        | Yes                                 |
 
 
 # Installation

diff --git a/mcli/mcli-1b-eval.yaml b/mcli/mcli-1b-eval.yaml
@@ -9,7 +9,7 @@ integrations:
 command: |
   cd llm-foundry/scripts/
   composer eval/eval.py /mnt/config/parameters.yaml
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 name: mpt-1b-eval
 
 compute:

diff --git a/mcli/mcli-1b-max-seq-len-8k.yaml b/mcli/mcli-1b-max-seq-len-8k.yaml
@@ -17,7 +17,7 @@ command: |
     --out_root ./my-copy-c4 --splits train_small val_small \
     --concat_tokens 8192 --tokenizer EleutherAI/gpt-neox-20b --eos_text '<|endoftext|>'
   composer train/train.py /mnt/config/parameters.yaml
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 name: mpt-1b-ctx-8k-gpus-8
 
 compute:

diff --git a/mcli/mcli-1b.yaml b/mcli/mcli-1b.yaml
@@ -21,7 +21,7 @@ command: |
     eval_loader.dataset.split=val_small \
     max_duration=100ba \
     eval_interval=0
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 name: mpt-1b-gpus-8
 
 compute:

diff --git a/mcli/mcli-benchmark-mpt.yaml b/mcli/mcli-benchmark-mpt.yaml
@@ -6,7 +6,7 @@ compute:
   # cluster: TODO # Name of the cluster to use for this run
   # gpu_type: a100_80gb # Type of GPU to use. We use a100_80gb in our experiments
 
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 
 integrations:
 - integration_type: git_repo

diff --git a/mcli/mcli-convert-composer-to-hf.yaml b/mcli/mcli-convert-composer-to-hf.yaml
@@ -13,7 +13,7 @@ command: |
     --hf_output_path s3://bucket/folder/hf/ \
     --output_precision bf16 \
 
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 name: convert-composer-hf
 
 compute:

diff --git a/mcli/mcli-hf-eval.yaml b/mcli/mcli-hf-eval.yaml
@@ -16,7 +16,7 @@ gpu_num: 8
 # gpu_type:
 # cluster:  # replace with your cluster here!
 
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 
 # The below is injected as a YAML file: /mnt/config/parameters.yaml
 parameters:

diff --git a/mcli/mcli-hf-generate.yaml b/mcli/mcli-hf-generate.yaml
@@ -35,7 +35,7 @@ command: |
       "Here's a quick recipe for baking chocolate chip cookies: Start by" \
       "The best 5 cities to visit in Europe are"
 
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 name: hf-generate
 
 compute:

diff --git a/mcli/mcli-llama2-finetune.yaml b/mcli/mcli-llama2-finetune.yaml
@@ -9,7 +9,7 @@ integrations:
 command: |
   cd llm-foundry/scripts
   composer train/train.py /mnt/config/parameters.yaml
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 name: llama2-finetune
 
 compute:

diff --git a/mcli/mcli-openai-eval.yaml b/mcli/mcli-openai-eval.yaml
@@ -16,7 +16,7 @@ gpu_num:  #
 gpu_type:  #
 cluster:  # replace with your cluster here!
 
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 
 # The below is injected as a YAML file: /mnt/config/parameters.yaml
 parameters:

diff --git a/mcli/mcli-pretokenize-oci-upload.yaml b/mcli/mcli-pretokenize-oci-upload.yaml
@@ -1,5 +1,5 @@
 name: c4-2k-pre-tokenized
-image: mosaicml/llm-foundry:2.2.1_cu121_flash2-latest
+image: mosaicml/llm-foundry:2.3.0_cu121_flash2-latest
 compute:
   gpus: 8  # Number of GPUs to use
 

diff --git a/setup.py b/setup.py
@@ -58,7 +58,7 @@
     'accelerate>=0.25,<0.26',  # for HF inference `device_map`
     'transformers>=4.40,<4.41',
     'mosaicml-streaming>=0.7.5,<0.8',
-    'torch>=2.2.1,<2.4',
+    'torch>=2.3.0,<2.4',
     'datasets>=2.19,<2.20',
     'fsspec==2023.6.0',  # newer version results in a bug in datasets that duplicates data
     'sentencepiece==0.1.97',
@@ -69,7 +69,7 @@
     'onnx==1.14.0',
     'onnxruntime==1.15.1',
     'boto3>=1.21.45,<2',
-    'huggingface-hub>=0.19.0,<1.0',
+    'huggingface-hub>=0.19.0,<0.23',
     'beautifulsoup4>=4.12.2,<5',  # required for model download utils
     'tenacity>=8.2.3,<9',
     'catalogue>=2,<3',

diff --git a/tests/a_scripts/inference/test_convert_composer_to_hf.py b/tests/a_scripts/inference/test_convert_composer_to_hf.py
@@ -790,6 +790,7 @@ def test_huggingface_conversion_callback(
     trainer = Trainer(
         model=original_model,
         device='gpu',
+        precision='amp_bf16',
         fsdp_config=fsdp_config if fsdp_state_dict_type is not None else None,
         train_dataloader=train_dataloader,
         save_folder=os.path.join(tmp_path, 'checkpoints'),

diff --git a/tests/data/test_dataloader.py b/tests/data/test_dataloader.py
@@ -385,6 +385,8 @@ def test_finetuning_dataloader(use_chat_formatting: bool,
 def test_finetuning_dataloader_safe_load(hf_name: str,
                                          hf_revision: Optional[str],
                                          expectation: ContextManager):
+    # Clear the folder
+    shutil.rmtree(DOWNLOADED_FT_DATASETS_DIRPATH, ignore_errors=True)
     cfg = DictConfig({
         'name': 'finetuning',
         'dataset': {

diff --git a/tests/eval/local_data/gsm8k_small.jsonl b/tests/eval/local_data/gsm8k_small.jsonl
@@ -1,4 +1,4 @@
 {"context": "Janet’s ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?", "chain_of_thought": "Janet sells 16 - 3 - 4 = <<16-3-4=9>>9 duck eggs a day.\nShe makes 9 * 2 = $<<9*2=18>>18 every day at the farmer’s market.", "answer": "18"}
-{"context": "A robe takes 2 bolts of blue fiber and half that much white fiber.  How many bolts in total does it take?", "chain_of_thought": "It takes 2/2=<<2/2=1>>1 bolt of white fiber\nSo the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric", "answer": "3"}
+{"context": "A robe takes 2 bolts of blue fiber and half that much white fiber.  How many bolts in total does it take?", "chain_of_thought": "It takes 2/2=<<2/2=1>>1 bolt of white fiber\nSo the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric", "answer": "12334"}
 {"context": "Josh decides to try flipping a house.  He buys a house for $80,000 and then puts in $50,000 in repairs.  This increased the value of the house by 150%.  How much profit did he make?", "chain_of_thought": "The cost of the house and repairs came out to 80,000+50,000=$<<80000+50000=130000>>130,000\nHe increased the value of the house by 80,000*1.5=<<80000*1.5=120000>>120,000\nSo the new value of the house is 120,000+80,000=$<<120000+80000=200000>>200,000\nSo he made a profit of 200,000-130,000=$<<200000-130000=70000>>70,000", "answer": "70000"}
 {"context": "James decides to run 3 sprints 3 times a week.  He runs 60 meters each sprint.  How many total meters does he run a week?", "chain_of_thought": "He sprints 3*3=<<3*3=9>>9 times\nSo he runs 9*60=<<9*60=540>>540 meters", "answer": "540"}
diff --git a/tests/eval/test_in_context_learning_datasets.py b/tests/eval/test_in_context_learning_datasets.py
@@ -1571,7 +1571,7 @@ def test_qa_task_with_cot_dataloader(
     assert all(item.count('Q: ') == num_fewshot + 1 for item in decoded_batch)
     assert all(item.count('\nA:') == num_fewshot + 1 for item in decoded_batch)
 
-    assert batch['labels'] == [['18'], ['3']]
+    assert batch['labels'] == [['18'], ['12334']]
     if num_fewshot == 0:
         assert decoded_batch[0].endswith(
             "Q: Janet’s ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?\nA: Let's think step by step."