Dynamic slurm generation. #84
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,194 @@ | ||
| from datetime import datetime | ||
| from pathlib import Path | ||
| from typing import Any | ||
|
|
||
|
|
||
| VLLM_TASK_MAP = { | ||
| "LLM": "generate", | ||
| "VLM": "generate", | ||
| "Text_Embedding": "embed", | ||
| "Reward_Modeling": "reward", | ||
| } | ||
|
|
||
|
|
||
| class SlurmScriptGenerator: | ||
| def __init__(self, params: dict[str, Any], src_dir: str): | ||
| self.params = params | ||
| self.src_dir = src_dir | ||
| self.is_multinode = int(self.params["num_nodes"]) > 1 | ||
| self.model_weights_path = str( | ||
| Path(params["model_weights_parent_dir"], params["model_name"]) | ||
| ) | ||
| self.task = VLLM_TASK_MAP[self.params["model_type"]] | ||
|
|
||
| def _generate_script_content(self) -> str: | ||
| preamble = self._generate_preamble() | ||
| server = self._generate_server_script() | ||
| env_exports = self._export_parallel_vars() | ||
| launcher = self._generate_launcher() | ||
| args = self._generate_shared_args() | ||
| return preamble + server + env_exports + launcher + args | ||
|
|
||
| def _generate_preamble(self) -> str: | ||
| base = [ | ||
| "#!/bin/bash", | ||
| "#SBATCH --cpus-per-task=16", | ||
| "#SBATCH --mem=64G", | ||
| ] | ||
| if self.is_multinode: | ||
| base += [ | ||
| "#SBATCH --exclusive", | ||
| "#SBATCH --tasks-per-node=1", | ||
| ] | ||
| base += [""] | ||
| return "\n".join(base) | ||
|
|
||
| def _export_parallel_vars(self) -> str: | ||
|
||
| if self.is_multinode: | ||
| return """if [ "$PIPELINE_PARALLELISM" = "True" ]; then | ||
|
||
| export PIPELINE_PARALLEL_SIZE=$SLURM_JOB_NUM_NODES | ||
| export TENSOR_PARALLEL_SIZE=$SLURM_GPUS_PER_NODE | ||
| else | ||
| export PIPELINE_PARALLEL_SIZE=1 | ||
| export TENSOR_PARALLEL_SIZE=$((SLURM_JOB_NUM_NODES*SLURM_GPUS_PER_NODE)) | ||
| fi | ||
|
|
||
| """ | ||
| return "export TENSOR_PARALLEL_SIZE=$SLURM_GPUS_PER_NODE\n\n" | ||
|
|
||
| def _generate_shared_args(self) -> str: | ||
| args = [ | ||
| f"--model {self.model_weights_path} \\", | ||
| f"--served-model-name {self.params['model_name']} \\", | ||
| '--host "0.0.0.0" \\', | ||
| "--port $vllm_port_number \\", | ||
| "--tensor-parallel-size ${TENSOR_PARALLEL_SIZE} \\", | ||
| f"--dtype {self.params['data_type']} \\", | ||
| "--trust-remote-code \\", | ||
| f"--max-logprobs {self.params['vocab_size']} \\", | ||
| f"--max-model-len {self.params['max_model_len']} \\", | ||
| f"--max-num-seqs {self.params['max_num_seqs']} \\", | ||
| f"--gpu-memory-utilization {self.params['gpu_memory_utilization']} \\", | ||
| f"--compilation-config {self.params['compilation_config']} \\", | ||
| f"--task {self.task} \\", | ||
| ] | ||
| if self.is_multinode: | ||
| args.insert(4, "--pipeline-parallel-size ${PIPELINE_PARALLEL_SIZE} \\") | ||
| if self.params.get("max_num_batched_tokens"): | ||
| args.append( | ||
| f"--max-num-batched-tokens={self.params['max_num_batched_tokens']} \\" | ||
| ) | ||
| if self.params.get("enable_prefix_caching") == "True": | ||
| args.append("--enable-prefix-caching \\") | ||
| if self.params.get("enable_chunked_prefill") == "True": | ||
| args.append("--enable-chunked-prefill \\") | ||
| if self.params.get("enforce_eager") == "True": | ||
| args.append("--enforce-eager") | ||
|
|
||
| return "\n".join(args) | ||
|
|
||
| def _generate_server_script(self) -> str: | ||
| server_script = [""] | ||
| if self.params["venv"] == "singularity": | ||
| server_script.append("""export SINGULARITY_IMAGE=/model-weights/vec-inf-shared/vector-inference_latest.sif | ||
| export VLLM_NCCL_SO_PATH=/vec-inf/nccl/libnccl.so.2.18.1 | ||
| module load singularity-ce/3.8.2 | ||
| singularity exec $SINGULARITY_IMAGE ray stop | ||
| """) | ||
| server_script.append(f"source {self.src_dir}/find_port.sh\n") | ||
| server_script.append( | ||
| self._generate_multinode_server_script() | ||
| if self.is_multinode | ||
| else self._generate_single_node_server_script() | ||
| ) | ||
| server_script.append(f"""echo "Updating server address in $JSON_PATH" | ||
| JSON_PATH="{self.params["log_dir"]}/{self.params["model_name"]}.$SLURM_JOB_ID/{self.params["model_name"]}.$SLURM_JOB_ID.json" | ||
| jq --arg server_addr "$SERVER_ADDR" \\ | ||
| '. + {{"server_address": $server_addr}}' \\ | ||
| "$JSON_PATH" > temp.json \\ | ||
| && mv temp.json "$JSON_PATH" \\ | ||
| && rm -f temp.json | ||
|
|
||
| """) | ||
| return "\n".join(server_script) | ||
|
|
||
| def _generate_single_node_server_script(self) -> str: | ||
| return """hostname=${SLURMD_NODENAME} | ||
| vllm_port_number=$(find_available_port ${hostname} 8080 65535) | ||
|
|
||
| SERVER_ADDR="http://${hostname}:${vllm_port_number}/v1" | ||
| echo "Server address: $SERVER_ADDR" | ||
| """ | ||
|
|
||
| def _generate_multinode_server_script(self) -> str: | ||
| server_script = [] | ||
| server_script.append("""nodes=$(scontrol show hostnames "$SLURM_JOB_NODELIST") | ||
| nodes_array=($nodes) | ||
|
|
||
| head_node=${nodes_array[0]} | ||
| head_node_ip=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address) | ||
|
|
||
| head_node_port=$(find_available_port $head_node_ip 8080 65535) | ||
|
|
||
| ip_head=$head_node_ip:$head_node_port | ||
| export ip_head | ||
| echo "IP Head: $ip_head" | ||
|
|
||
| echo "Starting HEAD at $head_node" | ||
| srun --nodes=1 --ntasks=1 -w "$head_node" \\""") | ||
|
|
||
| if self.params["venv"] == "singularity": | ||
| server_script.append( | ||
| f" singularity exec --nv --bind {self.model_weights_path}:{self.model_weights_path} $SINGULARITY_IMAGE \\" | ||
| ) | ||
|
|
||
| server_script.append(""" ray start --head --node-ip-address="$head_node_ip" --port=$head_node_port \\ | ||
| --num-cpus "${SLURM_CPUS_PER_TASK}" --num-gpus "${SLURM_GPUS_PER_NODE}" --block & | ||
|
|
||
| sleep 10 | ||
| worker_num=$((SLURM_JOB_NUM_NODES - 1)) | ||
|
|
||
| for ((i = 1; i <= worker_num; i++)); do | ||
| node_i=${nodes_array[$i]} | ||
| echo "Starting WORKER $i at $node_i" | ||
| srun --nodes=1 --ntasks=1 -w "$node_i" \\""") | ||
|
|
||
| if self.params["venv"] == "singularity": | ||
| server_script.append( | ||
| f""" singularity exec --nv --bind {self.model_weights_path}:{self.model_weights_path} $SINGULARITY_IMAGE \\""" | ||
| ) | ||
| server_script.append(""" ray start --address "$ip_head" \\ | ||
| --num-cpus "${SLURM_CPUS_PER_TASK}" --num-gpus "${SLURM_GPUS_PER_NODE}" --block & | ||
| sleep 5 | ||
| done | ||
|
|
||
| vllm_port_number=$(find_available_port $head_node_ip 8080 65535) | ||
|
|
||
| SERVER_ADDR="http://${head_node_ip}:${vllm_port_number}/v1" | ||
| echo "Server address: $SERVER_ADDR" | ||
|
|
||
| """) | ||
| return "\n".join(server_script) | ||
|
|
||
| def _generate_launcher(self) -> str: | ||
| if self.params["venv"] == "singularity": | ||
| launcher_script = [ | ||
| f"""singularity exec --nv --bind {self.model_weights_path}:{self.model_weights_path} $SINGULARITY_IMAGE \\""" | ||
| ] | ||
| else: | ||
| launcher_script = [f"""source {self.params["venv"]}/bin/activate"""] | ||
| launcher_script.append( | ||
| """python3.10 -m vllm.entrypoints.openai.api_server \\\n""" | ||
| ) | ||
| return "\n".join(launcher_script) | ||
|
|
||
| def write_to_log_dir(self) -> Path: | ||
| log_subdir: Path = Path(self.params["log_dir"]) / self.params["model_name"] | ||
| log_subdir.mkdir(parents=True, exist_ok=True) | ||
|
|
||
| timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") | ||
| script_path: Path = log_subdir / f"launch_{timestamp}.slurm" | ||
|
|
||
| content = self._generate_script_content() | ||
| script_path.write_text(content) | ||
| return script_path | ||
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Don't think we need this
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, but when I didn't have it, it was trying to use Python 3.12 and ran into dependency issues. That's why I changed it. But I guess we can leave that for another PR if it's still a problem.
I was still using Poetry.