ob-metaflow-extensions 1.1.158__py2.py3-none-any.whl → 1.1.159__py2.py3-none-any.whl

metaflow_extensions/outerbounds/plugins/__init__.py

@@ -347,4 +347,4 @@ SECRETS_PROVIDERS_DESC = [
     ("outerbounds", ".secrets.secrets.OuterboundsSecretsProvider"),
 ]
 # Adding an override here so the library can be imported at the metaflow.plugins level
-__mf_promote_submodules__ = ["snowflake", "ollama"]
+__mf_promote_submodules__ = ["snowflake", "ollama", "torchtune"]

metaflow_extensions/outerbounds/plugins/torchtune/__init__.py

@@ -0,0 +1,159 @@
+from queue import Queue, Empty
+from concurrent.futures import ThreadPoolExecutor
+from typing import Optional, List, Dict
+import subprocess
+import shutil
+import sys
+from metaflow import current
+
+__mf_promote_submodules__ = ["plugins.torchtune"]
+
+
+class TorchTune:
+    def __init__(
+        self,
+        use_multi_node_config: bool = False,
+    ):
+        """
+        Initialize the Tune launcher.
+
+        :param use_multi_node_config: If True, attempt to build a distributed configuration
+                                      from current.torch.torchrun_args.
+        """
+        self.multi_node_config = {}
+        if use_multi_node_config:
+            if getattr(current, "torch", None):
+                print(
+                    "[Metaflow Tune] Since @torchrun is used, multi-node config can be used to launch the job."
+                )
+                # For distributed torchtune launches, we use similar parameters as torchrun.
+                # (You might need to adjust the keys according to your environment.)
+                self.multi_node_config = {
+                    "nnodes": current.torch.torchrun_args["nnodes"],
+                    "master_addr": current.torch.torchrun_args["master_addr"],
+                    "master_port": int(current.torch.torchrun_args["master_port"]),
+                    "node_rank": current.torch.torchrun_args["node_rank"],
+                    "nproc_per_node": current.torch.torchrun_args["nproc_per_node"],
+                    "num_processes": current.torch.torchrun_args["nproc_per_node"]
+                    * current.torch.torchrun_args["nnodes"],
+                }
+                print(
+                    f"[Metaflow Tune] Discovered multi-node config for torchrun: {self.multi_node_config}"
+                )
+            else:
+                print(
+                    "[Metaflow Tune] Since @torchrun is not used, default multi-node config cannot be used to launch the job."
+                )
+
+    def run(
+        self,
+        recipe: str,
+        config_dict: Dict,
+        additional_cli_options: Optional[List[str]] = None,
+    ):
+        """
+        Launch the torchtune job via its CLI.
+
+        :param recipe: The path to the recipe (or name of the recipe) to run.
+        :param config_dict: Optional dictionary that will be dumped to a YAML file and passed via --config.
+        :param additional_cli_options: Optional list of additional CLI options.
+        :raises: subprocess.CalledProcessError if the subprocess returns a nonzero exit code.
+        """
+        import yaml
+        import tempfile
+        import os
+
+        _temp_dir = tempfile.mkdtemp()
+        try:
+            config_path = os.path.join(_temp_dir, "config.yaml")
+            with open(config_path, "w") as f:
+                yaml.dump(config_dict, f)
+
+            additional_options = (
+                additional_cli_options if additional_cli_options else []
+            )
+
+            # Build the command. Here we use "tune run" as the base command.
+            cmd = ["tune", "run"]
+
+            # If distributed configuration is present, add torchrun–style flags.
+            if self.multi_node_config:
+                cmd.extend(
+                    [
+                        "--nnodes",
+                        str(self.multi_node_config.get("nnodes")),
+                        "--nproc-per-node",
+                        str(self.multi_node_config.get("nproc_per_node")),
+                        # "--rdzv_conf", f"rdzv_endpoint={self.multi_node_config.get('master_addr')}:{self.multi_node_config.get('master_port')}"
+                        "--rdzv-backend",
+                        "c10d",
+                        "--rdzv-endpoint",
+                        f"{self.multi_node_config.get('master_addr')}:{self.multi_node_config.get('master_port')}",
+                        "--rdzv-id",
+                        "1234567890",
+                        "--node-rank",
+                        str(self.multi_node_config.get("node_rank")),
+                        # TODO: should there be a masterip/port here ?
+                    ]
+                )
+
+            cmd.extend(additional_options)
+
+            cmd.append(recipe)
+            # If a recipe configuration was provided, pass it via the --config flag.
+            cmd.extend(["--config", config_path])
+
+            # Append any additional CLI options.
+
+            # Launch the subprocess.
+            print(f"[Metaflow tune] {' '.join(cmd)}")
+            process = subprocess.Popen(
+                cmd,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                universal_newlines=True,
+            )
+
+            # Stream the output in real-time.
+            for out_line, err_line in read_popen_pipes(process):
+                print(out_line, end="", flush=True)
+                print(err_line, end="", file=sys.stderr, flush=True)
+
+            process.wait()
+            if process.returncode != 0:
+                raise subprocess.CalledProcessError(process.returncode, cmd)
+        finally:
+            shutil.rmtree(_temp_dir)
+
+
+def enqueue_output(file, queue):
+    for line in iter(file.readline, ""):
+        queue.put(line)
+    file.close()
+
+
+def read_popen_pipes(p):
+
+    with ThreadPoolExecutor(2) as pool:
+        q_stdout, q_stderr = Queue(), Queue()
+
+        pool.submit(enqueue_output, p.stdout, q_stdout)
+        pool.submit(enqueue_output, p.stderr, q_stderr)
+
+        while True:
+
+            if p.poll() is not None and q_stdout.empty() and q_stderr.empty():
+                break
+
+            out_line = err_line = ""
+
+            try:
+                out_line = q_stdout.get_nowait()
+            except Empty:
+                pass
+            try:
+                err_line = q_stderr.get_nowait()
+            except Empty:
+                pass
+
+            yield (out_line, err_line)

metaflow_extensions/outerbounds/toplevel/plugins/torchtune/__init__.py

@@ -0,0 +1 @@
+__mf_promote_submodules__ = ["plugins.torchtune"]

{ob_metaflow_extensions-1.1.158.dist-info → ob_metaflow_extensions-1.1.159.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ob-metaflow-extensions
-Version: 1.1.158
+Version: 1.1.159
 Summary: Outerbounds Platform Extensions for Metaflow
 Author: Outerbounds, Inc.
 License: Commercial

{ob_metaflow_extensions-1.1.158.dist-info → ob_metaflow_extensions-1.1.159.dist-info}/RECORD

@@ -1,7 +1,7 @@
 metaflow_extensions/outerbounds/__init__.py,sha256=Gb8u06s9ClQsA_vzxmkCzuMnigPy7kKcDnLfb7eB-64,514
 metaflow_extensions/outerbounds/remote_config.py,sha256=pEFJuKDYs98eoB_-ryPjVi9b_c4gpHMdBHE14ltoxIU,4672
 metaflow_extensions/outerbounds/config/__init__.py,sha256=JsQGRuGFz28fQWjUvxUgR8EKBLGRdLUIk_buPLJplJY,1225
-metaflow_extensions/outerbounds/plugins/__init__.py,sha256=gytuNt3lNabirHLEYzrmHFMp-JWh8dA2AZPK11HmaNw,13242
+metaflow_extensions/outerbounds/plugins/__init__.py,sha256=GxYKjrMJCGVKoxhfdPAlVF9kYrEb3-xn9fgUTb_H9VY,13255
 metaflow_extensions/outerbounds/plugins/auth_server.py,sha256=_Q9_2EL0Xy77bCRphkwT1aSu8gQXRDOH-Z-RxTUO8N4,2202
 metaflow_extensions/outerbounds/plugins/perimeters.py,sha256=QXh3SFP7GQbS-RAIxUOPbhPzQ7KDFVxZkTdKqFKgXjI,2697
 metaflow_extensions/outerbounds/plugins/apps/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -61,6 +61,7 @@ metaflow_extensions/outerbounds/plugins/snowpark/snowpark_exceptions.py,sha256=F
 metaflow_extensions/outerbounds/plugins/snowpark/snowpark_job.py,sha256=aQphxX6jqYgfa83w387pEWl0keuLm38V53I8P8UL2ck,6887
 metaflow_extensions/outerbounds/plugins/snowpark/snowpark_service_spec.py,sha256=AI_kcm1hZV3JRxJkookcH6twiGnAYjk9Dx-MeoYz60Y,8511
 metaflow_extensions/outerbounds/plugins/tensorboard/__init__.py,sha256=9lUM4Cqi5RjrHBRfG6AQMRz8-R96eZC8Ih0KD2lv22Y,1858
+metaflow_extensions/outerbounds/plugins/torchtune/__init__.py,sha256=TOXNeyhcgd8VxplXO_oEuryFEsbk0tikn5GL0-44SU8,5853
 metaflow_extensions/outerbounds/profilers/__init__.py,sha256=wa_jhnCBr82TBxoS0e8b6_6sLyZX0fdHicuGJZNTqKw,29
 metaflow_extensions/outerbounds/profilers/gpu.py,sha256=3Er8uKQzfm_082uadg4yn_D4Y-iSCgzUfFmguYxZsz4,27485
 metaflow_extensions/outerbounds/toplevel/__init__.py,sha256=qWUJSv_r5hXJ7jV_On4nEasKIfUCm6_UjkjXWA_A1Ts,90
@@ -70,7 +71,8 @@ metaflow_extensions/outerbounds/toplevel/plugins/gcp/__init__.py,sha256=BbZiaH3u
 metaflow_extensions/outerbounds/toplevel/plugins/kubernetes/__init__.py,sha256=5zG8gShSj8m7rgF4xgWBZFuY3GDP5n1T0ktjRpGJLHA,69
 metaflow_extensions/outerbounds/toplevel/plugins/ollama/__init__.py,sha256=GRSz2zwqkvlmFS6bcfYD_CX6CMko9DHQokMaH1iBshA,47
 metaflow_extensions/outerbounds/toplevel/plugins/snowflake/__init__.py,sha256=LptpH-ziXHrednMYUjIaosS1SXD3sOtF_9_eRqd8SJw,50
-ob_metaflow_extensions-1.1.158.dist-info/METADATA,sha256=0t_P8-Uhi3I39xyeSGv2BpRQO5Upe1eIjs04e6Stjd8,521
-ob_metaflow_extensions-1.1.158.dist-info/WHEEL,sha256=bb2Ot9scclHKMOLDEHY6B2sicWOgugjFKaJsT7vwMQo,110
-ob_metaflow_extensions-1.1.158.dist-info/top_level.txt,sha256=NwG0ukwjygtanDETyp_BUdtYtqIA_lOjzFFh1TsnxvI,20
-ob_metaflow_extensions-1.1.158.dist-info/RECORD,,
+metaflow_extensions/outerbounds/toplevel/plugins/torchtune/__init__.py,sha256=uTVkdSk3xZ7hEKYfdlyVteWj5KeDwaM1hU9WT-_YKfI,50
+ob_metaflow_extensions-1.1.159.dist-info/METADATA,sha256=B88tVfHuAktlRjsOBy8-pwquOYyBJUuOctDcL2KfN9I,521
+ob_metaflow_extensions-1.1.159.dist-info/WHEEL,sha256=bb2Ot9scclHKMOLDEHY6B2sicWOgugjFKaJsT7vwMQo,110
+ob_metaflow_extensions-1.1.159.dist-info/top_level.txt,sha256=NwG0ukwjygtanDETyp_BUdtYtqIA_lOjzFFh1TsnxvI,20
+ob_metaflow_extensions-1.1.159.dist-info/RECORD,,