[tutorial] edited hands-on practices (#1899)

* Add handson to ColossalAI.

* Change names of handsons and edit sequence parallel example.

* Edit wrong folder name

* resolve conflict

* delete readme

examples/tutorial/hybrid_parallel/README.md

+# Handson 1: Multi-dimensional Parallelism with Colossal-AI
+
+
+## Install Colossal-AI and other dependencies
+
+```bash
+sh install.sh
+```
+
+
+## Prepare Dataset
+
+We use CIFAR10 dataset in this example. The dataset will be downloaded to `../data` by default. 
+If you wish to use customized directory for the dataset. You can set the environment variable `DATA` via the following command.
+
+```bash
+export DATA=/path/to/data
+```
+
+
+## Run on 2*2 device mesh
+
+Current configuration setting on `config.py` is TP=2, PP=2.
+
+```bash
+colossalai run --nproc_per_node 4 train.py --config config.py
+```
\ No newline at end of file

examples/tutorial/hybrid_parallel/config.py

+from colossalai.amp import AMP_TYPE
+
+# hyperparameters
+# BATCH_SIZE is as per GPU
+# global batch size = BATCH_SIZE x data parallel size
+BATCH_SIZE = 256
+LEARNING_RATE = 3e-3
+WEIGHT_DECAY = 0.3
+NUM_EPOCHS = 10
+WARMUP_EPOCHS = 3
+
+# model config
+IMG_SIZE = 224
+PATCH_SIZE = 16
+HIDDEN_SIZE = 512
+DEPTH = 4
+NUM_HEADS = 4
+MLP_RATIO = 2
+NUM_CLASSES = 1000
+CHECKPOINT = False
+SEQ_LENGTH = (IMG_SIZE // PATCH_SIZE)**2 + 1    # add 1 for cls token
+
+# parallel setting
+TENSOR_PARALLEL_SIZE = 2
+TENSOR_PARALLEL_MODE = '1d'
+
+parallel = dict(
+    pipeline=2,
+    tensor=dict(mode=TENSOR_PARALLEL_MODE, size=TENSOR_PARALLEL_SIZE),
+)
+
+fp16 = dict(mode=AMP_TYPE.NAIVE)
+clip_grad_norm = 1.0
+
+# pipeline config
+NUM_MICRO_BATCHES = parallel['pipeline']

examples/tutorial/hybrid_parallel/install.sh

+pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
+pip install colossalai==0.1.10+torch1.12cu11.3 -f https://release.colossalai.org
+pip install titans
+colossalai check -i
\ No newline at end of file

examples/tutorial/hybrid_parallel/train.py

+import os
+import colossalai
+import torch
+
+from tqdm import tqdm
+from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
+from colossalai.logging import get_dist_logger
+from colossalai.nn import CrossEntropyLoss
+from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
+from colossalai.utils import is_using_pp, get_dataloader
+from colossalai.pipeline.pipelinable import PipelinableContext
+from titans.model.vit.vit import _create_vit_model
+from titans.dataloader.cifar10 import build_cifar
+
+
+def main():
+    # initialize distributed setting
+    parser = colossalai.get_default_parser()
+    args = parser.parse_args()
+
+    # launch from torch
+    colossalai.launch_from_torch(config=args.config)
+
+    # get logger
+    logger = get_dist_logger()
+    logger.info("initialized distributed environment", ranks=[0])
+
+    if hasattr(gpc.config, 'LOG_PATH'):
+        if gpc.get_global_rank() == 0:
+            log_path = gpc.config.LOG_PATH
+            if not os.path.exists(log_path):
+                os.mkdir(log_path)
+            logger.log_to_file(log_path)
+
+    use_pipeline = is_using_pp()
+
+    # create model
+    model_kwargs = dict(img_size=gpc.config.IMG_SIZE,
+                        patch_size=gpc.config.PATCH_SIZE,
+                        hidden_size=gpc.config.HIDDEN_SIZE,
+                        depth=gpc.config.DEPTH,
+                        num_heads=gpc.config.NUM_HEADS,
+                        mlp_ratio=gpc.config.MLP_RATIO,
+                        num_classes=10,
+                        init_method='jax',
+                        checkpoint=gpc.config.CHECKPOINT)
+
+    if use_pipeline:
+        pipelinable = PipelinableContext()
+        with pipelinable:
+            model = _create_vit_model(**model_kwargs)
+        pipelinable.to_layer_list()
+        pipelinable.policy = "uniform"
+        model = pipelinable.partition(
+            1, gpc.pipeline_parallel_size, gpc.get_local_rank(ParallelMode.PIPELINE))
+    else:
+        model = _create_vit_model(**model_kwargs)
+
+    # count number of parameters
+    total_numel = 0
+    for p in model.parameters():
+        total_numel += p.numel()
+    if not gpc.is_initialized(ParallelMode.PIPELINE):
+        pipeline_stage = 0
+    else:
+        pipeline_stage = gpc.get_local_rank(ParallelMode.PIPELINE)
+    logger.info(
+        f"number of parameters: {total_numel} on pipeline stage {pipeline_stage}")
+
+    # create dataloaders
+    root = os.environ.get('DATA', '../data/cifar10')
+    train_dataloader, test_dataloader = build_cifar(
+        gpc.config.BATCH_SIZE, root, pad_if_needed=True)
+
+    # create loss function
+    criterion = CrossEntropyLoss(label_smoothing=0.1)
+
+    # create optimizer
+    optimizer = torch.optim.AdamW(model.parameters(
+    ), lr=gpc.config.LEARNING_RATE, weight_decay=gpc.config.WEIGHT_DECAY)
+
+    # create lr scheduler
+    lr_scheduler = CosineAnnealingWarmupLR(optimizer=optimizer,
+                                           total_steps=gpc.config.NUM_EPOCHS,
+                                           warmup_steps=gpc.config.WARMUP_EPOCHS)
+
+    # initialize
+    engine, train_dataloader, test_dataloader, _ = colossalai.initialize(model=model,
+                                                                         optimizer=optimizer,
+                                                                         criterion=criterion,
+                                                                         train_dataloader=train_dataloader,
+                                                                         test_dataloader=test_dataloader)
+
+    logger.info("Engine is built", ranks=[0])
+
+    data_iter = iter(train_dataloader)
+
+    for epoch in range(gpc.config.NUM_EPOCHS):
+        # training
+        engine.train()
+
+        if gpc.get_global_rank() == 0:
+            description = 'Epoch {} / {}'.format(epoch, gpc.config.NUM_EPOCHS)
+            progress = tqdm(range(len(train_dataloader)), desc=description)
+        else:
+            progress = range(len(train_dataloader))
+        for _ in progress:
+            engine.zero_grad()
+            engine.execute_schedule(data_iter, return_output_label=False)
+            engine.step()
+            lr_scheduler.step()
+
+
+if __name__ == '__main__':
+    main()

examples/tutorial/large_batch_optimizer/README.md

+# Handson 4: Comparison of Large Batch Training Optimization
+
+## Prepare Dataset
+
+We use CIFAR10 dataset in this example. The dataset will be downloaded to `../data` by default. 
+If you wish to use customized directory for the dataset. You can set the environment variable `DATA` via the following command.
+
+```bash
+export DATA=/path/to/data
+```
+
+
+## Run on 2*2 device mesh
+
+```bash
+colossalai run --nproc_per_node 4 train.py --config config.py
+```
\ No newline at end of file

examples/tutorial/large_batch_optimizer/config.py

+from colossalai.amp import AMP_TYPE
+
+# hyperparameters
+# BATCH_SIZE is as per GPU
+# global batch size = BATCH_SIZE x data parallel size
+BATCH_SIZE = 512
+LEARNING_RATE = 3e-3
+WEIGHT_DECAY = 0.3
+NUM_EPOCHS = 10
+WARMUP_EPOCHS = 3
+
+# model config
+IMG_SIZE = 224
+PATCH_SIZE = 16
+HIDDEN_SIZE = 512
+DEPTH = 4
+NUM_HEADS = 4
+MLP_RATIO = 2
+NUM_CLASSES = 1000
+CHECKPOINT = False
+SEQ_LENGTH = (IMG_SIZE // PATCH_SIZE)**2 + 1    # add 1 for cls token
+
+# parallel setting
+TENSOR_PARALLEL_SIZE = 2
+TENSOR_PARALLEL_MODE = '1d'
+
+parallel = dict(
+    pipeline=2,
+    tensor=dict(mode=TENSOR_PARALLEL_MODE, size=TENSOR_PARALLEL_SIZE),
+)
+
+fp16 = dict(mode=AMP_TYPE.NAIVE)
+clip_grad_norm = 1.0
+
+# pipeline config
+NUM_MICRO_BATCHES = parallel['pipeline']

examples/tutorial/large_batch_optimizer/train.py

+import os
+import colossalai
+import torch
+
+from tqdm import tqdm
+from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
+from colossalai.logging import get_dist_logger
+from colossalai.nn import CrossEntropyLoss
+from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
+from colossalai.nn.optimizer import Lars, Lamb
+from colossalai.utils import is_using_pp, get_dataloader
+from colossalai.pipeline.pipelinable import PipelinableContext
+from titans.model.vit.vit import _create_vit_model
+from titans.dataloader.cifar10 import build_cifar
+
+
+def main():
+    # initialize distributed setting
+    parser = colossalai.get_default_parser()
+    args = parser.parse_args()
+
+    # launch from torch
+    colossalai.launch_from_torch(config=args.config)
+
+    # get logger
+    logger = get_dist_logger()
+    logger.info("initialized distributed environment", ranks=[0])
+
+    if hasattr(gpc.config, 'LOG_PATH'):
+        if gpc.get_global_rank() == 0:
+            log_path = gpc.config.LOG_PATH
+            if not os.path.exists(log_path):
+                os.mkdir(log_path)
+            logger.log_to_file(log_path)
+
+    use_pipeline = is_using_pp()
+
+    # create model
+    model_kwargs = dict(img_size=gpc.config.IMG_SIZE,
+                        patch_size=gpc.config.PATCH_SIZE,
+                        hidden_size=gpc.config.HIDDEN_SIZE,
+                        depth=gpc.config.DEPTH,
+                        num_heads=gpc.config.NUM_HEADS,
+                        mlp_ratio=gpc.config.MLP_RATIO,
+                        num_classes=10,
+                        init_method='jax',
+                        checkpoint=gpc.config.CHECKPOINT)
+
+    if use_pipeline:
+        pipelinable = PipelinableContext()
+        with pipelinable:
+            model = _create_vit_model(**model_kwargs)
+        pipelinable.to_layer_list()
+        pipelinable.policy = "uniform"
+        model = pipelinable.partition(
+            1, gpc.pipeline_parallel_size, gpc.get_local_rank(ParallelMode.PIPELINE))
+    else:
+        model = _create_vit_model(**model_kwargs)
+
+    # count number of parameters
+    total_numel = 0
+    for p in model.parameters():
+        total_numel += p.numel()
+    if not gpc.is_initialized(ParallelMode.PIPELINE):
+        pipeline_stage = 0
+    else:
+        pipeline_stage = gpc.get_local_rank(ParallelMode.PIPELINE)
+    logger.info(
+        f"number of parameters: {total_numel} on pipeline stage {pipeline_stage}")
+
+    # create dataloaders
+    root = os.environ.get('DATA', '../data/cifar10')
+    train_dataloader, test_dataloader = build_cifar(
+        gpc.config.BATCH_SIZE, root, pad_if_needed=True)
+
+    # create loss function
+    criterion = CrossEntropyLoss(label_smoothing=0.1)
+
+    # create optimizer
+    optimizer = Lars(model.parameters(), lr=gpc.config.LEARNING_RATE,
+                     weight_decay=gpc.config.WEIGHT_DECAY)
+
+    # create lr scheduler
+    lr_scheduler = CosineAnnealingWarmupLR(optimizer=optimizer,
+                                           total_steps=gpc.config.NUM_EPOCHS,
+                                           warmup_steps=gpc.config.WARMUP_EPOCHS)
+
+    # initialize
+    engine, train_dataloader, test_dataloader, _ = colossalai.initialize(model=model,
+                                                                         optimizer=optimizer,
+                                                                         criterion=criterion,
+                                                                         train_dataloader=train_dataloader,
+                                                                         test_dataloader=test_dataloader)
+
+    logger.info("Engine is built", ranks=[0])
+
+    data_iter = iter(train_dataloader)
+
+    for epoch in range(gpc.config.NUM_EPOCHS):
+        # training
+        engine.train()
+
+        if gpc.get_global_rank() == 0:
+            description = 'Epoch {} / {}'.format(epoch, gpc.config.NUM_EPOCHS)
+            progress = tqdm(range(len(train_dataloader)), desc=description)
+        else:
+            progress = range(len(train_dataloader))
+        for _ in progress:
+            engine.zero_grad()
+            engine.execute_schedule(data_iter, return_output_label=False)
+            engine.step()
+            lr_scheduler.step()
+
+
+if __name__ == '__main__':
+    main()

examples/tutorial/opt/README.md

+# Handson 5: Fine-tuning and Serving for OPT from Hugging Face

examples/tutorial/opt/inference/README.md

+# Overview
+
+This is an example showing how to run OPT generation. The OPT model is implemented using ColossalAI.
+
+It supports tensor parallelism, batching and caching.
+
+# How to run
+
+Run OPT-125M:
+```shell
+python opt_fastapi.py opt-125m
+```
+
+It will launch a HTTP server on `0.0.0.0:7070` by default and you can customize host and port. You can open `localhost:7070/docs` in your browser to see the openapi docs.
+
+## Configure
+
+### Configure model
+```shell
+python opt_fastapi.py <model>
+```
+Available models: opt-125m, opt-6.7b, opt-30b, opt-175b.
+
+### Configure tensor parallelism
+```shell
+python opt_fastapi.py <model> --tp <TensorParallelismWorldSize>
+```
+The `<TensorParallelismWorldSize>` can be an integer in `[1, #GPUs]`. Default `1`.
+
+### Configure checkpoint
+```shell
+python opt_fastapi.py <model> --checkpoint <CheckpointPath>
+```
+The `<CheckpointPath>` can be a file path or a directory path. If it's a directory path, all files under the directory will be loaded.
+
+### Configure queue
+```shell
+python opt_fastapi.py <model> --queue_size <QueueSize>
+```
+The `<QueueSize>` can be an integer in `[0, MAXINT]`. If it's `0`, the request queue size is infinite. If it's a positive integer, when the request queue is full, incoming requests will be dropped (the HTTP status code of response will be 406).
+
+### Configure bathcing
+```shell
+python opt_fastapi.py <model> --max_batch_size <MaxBatchSize>
+```
+The `<MaxBatchSize>` can be an integer in `[1, MAXINT]`. The engine will make batch whose size is less or equal to this value.
+
+Note that the batch size is not always equal to `<MaxBatchSize>`, as some consecutive requests may not be batched.
+
+### Configure caching
+```shell
+python opt_fastapi.py <model> --cache_size <CacheSize> --cache_list_size <CacheListSize>
+```
+This will cache `<CacheSize>` unique requests. And for each unique request, it cache `<CacheListSize>` different results. A random result will be returned if the cache is hit.
+
+The `<CacheSize>` can be an integer in `[0, MAXINT]`. If it's `0`, cache won't be applied. The `<CacheListSize>` can be an integer in `[1, MAXINT]`.
+
+### Other configurations
+```shell
+python opt_fastapi.py -h
+```
+
+# How to benchmark
+```shell
+cd benchmark
+locust
+```
+
+Then open the web interface link which is on your console.
+
+# Pre-process pre-trained weights
+
+## OPT-66B
+See [script/processing_ckpt_66b.py](./script/processing_ckpt_66b.py).
+
+## OPT-175B
+See [script/process-opt-175b](./script/process-opt-175b/).
\ No newline at end of file

examples/tutorial/opt/inference/batch.py

+import torch
+from typing import List, Deque, Tuple, Hashable, Any
+from energonai import BatchManager, SubmitEntry, TaskEntry
+
+
+class BatchManagerForGeneration(BatchManager):
+    def __init__(self, max_batch_size: int = 1, pad_token_id: int = 0) -> None:
+        super().__init__()
+        self.max_batch_size = max_batch_size
+        self.pad_token_id = pad_token_id
+
+    def _left_padding(self, batch_inputs):
+        max_len = max(len(inputs['input_ids']) for inputs in batch_inputs)
+        outputs = {'input_ids': [], 'attention_mask': []}
+        for inputs in batch_inputs:
+            input_ids, attention_mask = inputs['input_ids'], inputs['attention_mask']
+            padding_len = max_len - len(input_ids)
+            input_ids = [self.pad_token_id] * padding_len + input_ids
+            attention_mask = [0] * padding_len + attention_mask
+            outputs['input_ids'].append(input_ids)
+            outputs['attention_mask'].append(attention_mask)
+        for k in outputs:
+            outputs[k] = torch.tensor(outputs[k])
+        return outputs, max_len
+
+    @staticmethod
+    def _make_batch_key(entry: SubmitEntry) -> tuple:
+        data = entry.data
+        return (data['top_k'], data['top_p'], data['temperature'])
+
+    def make_batch(self, q: Deque[SubmitEntry]) -> Tuple[TaskEntry, dict]:
+        entry = q.popleft()
+        uids = [entry.uid]
+        batch = [entry.data]
+        while len(batch) < self.max_batch_size:
+            if len(q) == 0:
+                break
+            if self._make_batch_key(entry) != self._make_batch_key(q[0]):
+                break
+            if q[0].data['max_tokens'] > entry.data['max_tokens']:
+                break
+            e = q.popleft()
+            batch.append(e.data)
+            uids.append(e.uid)
+        inputs, max_len = self._left_padding(batch)
+        trunc_lens = []
+        for data in batch:
+            trunc_lens.append(max_len + data['max_tokens'])
+        inputs['top_k'] = entry.data['top_k']
+        inputs['top_p'] = entry.data['top_p']
+        inputs['temperature'] = entry.data['temperature']
+        inputs['max_tokens'] = max_len + entry.data['max_tokens']
+        return TaskEntry(tuple(uids), inputs), {'trunc_lens': trunc_lens}
+
+    def split_batch(self, task_entry: TaskEntry, trunc_lens: List[int] = []) -> List[Tuple[Hashable, Any]]:
+        retval = []
+        for uid, output, trunc_len in zip(task_entry.uids, task_entry.batch, trunc_lens):
+            retval.append((uid, output[:trunc_len]))
+        return retval

examples/tutorial/opt/inference/benchmark/locustfile.py

+from locust import HttpUser, task
+from json import JSONDecodeError
+
+
+class GenerationUser(HttpUser):
+    @task
+    def generate(self):
+        prompt = 'Question: What is the longest river on the earth? Answer:'
+        for i in range(4, 9):
+            data = {'max_tokens': 2**i, 'prompt': prompt}
+            with self.client.post('/generation', json=data, catch_response=True) as response:
+                if response.status_code in (200, 406):
+                    response.success()
+                else:
+                    response.failure('Response wrong')

examples/tutorial/opt/inference/cache.py

+from collections import OrderedDict
+from threading import Lock
+from contextlib import contextmanager
+from typing import List, Any, Hashable, Dict
+
+
+class MissCacheError(Exception):
+    pass
+
+
+class ListCache:
+    def __init__(self, cache_size: int, list_size: int, fixed_keys: List[Hashable] = []) -> None:
+        """Cache a list of values. The fixed keys won't be removed. For other keys, LRU is applied.
+        When the value list is not full, a cache miss occurs. Otherwise, a cache hit occurs. Redundant values will be removed.
+
+        Args:
+            cache_size (int): Max size for LRU cache.
+            list_size (int): Value list size.
+            fixed_keys (List[Hashable], optional): The keys which won't be removed. Defaults to [].
+        """
+        self.cache_size = cache_size
+        self.list_size = list_size
+        self.cache: OrderedDict[Hashable, List[Any]] = OrderedDict()
+        self.fixed_cache: Dict[Hashable, List[Any]] = {}
+        for key in fixed_keys:
+            self.fixed_cache[key] = []
+        self._lock = Lock()
+
+    def get(self, key: Hashable) -> List[Any]:
+        with self.lock():
+            if key in self.fixed_cache:
+                l = self.fixed_cache[key]
+                if len(l) >= self.list_size:
+                    return l
+            elif key in self.cache:
+                self.cache.move_to_end(key)
+                l = self.cache[key]
+                if len(l) >= self.list_size:
+                    return l
+        raise MissCacheError()
+
+    def add(self, key: Hashable, value: Any) -> None:
+        with self.lock():
+            if key in self.fixed_cache:
+                l = self.fixed_cache[key]
+                if len(l) < self.list_size and value not in l:
+                    l.append(value)
+            elif key in self.cache:
+                self.cache.move_to_end(key)
+                l = self.cache[key]
+                if len(l) < self.list_size and value not in l:
+                    l.append(value)
+            else:
+                if len(self.cache) >= self.cache_size:
+                    self.cache.popitem(last=False)
+                self.cache[key] = [value]
+
+    @contextmanager
+    def lock(self):
+        try:
+            self._lock.acquire()
+            yield
+        finally:
+            self._lock.release()

examples/tutorial/opt/inference/opt_fastapi.py

+import argparse
+import logging
+import random
+from typing import Optional
+
+import uvicorn
+from energonai import QueueFullError, launch_engine
+from energonai.model import opt_6B, opt_30B, opt_125M, opt_175B
+from fastapi import FastAPI, HTTPException, Request
+from pydantic import BaseModel, Field
+from transformers import GPT2Tokenizer
+
+from batch import BatchManagerForGeneration
+from cache import ListCache, MissCacheError
+
+
+class GenerationTaskReq(BaseModel):
+    max_tokens: int = Field(gt=0, le=256, example=64)
+    prompt: str = Field(
+        min_length=1, example='Question: Where were the 2004 Olympics held?\nAnswer: Athens, Greece\n\nQuestion: What is the longest river on the earth?\nAnswer:')
+    top_k: Optional[int] = Field(default=None, gt=0, example=50)
+    top_p: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.5)
+    temperature: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.7)
+
+
+app = FastAPI()
+
+
+@app.post('/generation')
+async def generate(data: GenerationTaskReq, request: Request):
+    logger.info(f'{request.client.host}:{request.client.port} - "{request.method} {request.url.path}" - {data}')
+    key = (data.prompt, data.max_tokens)
+    try:
+        if cache is None:
+            raise MissCacheError()
+        outputs = cache.get(key)
+        output = random.choice(outputs)
+        logger.info('Cache hit')
+    except MissCacheError:
+        inputs = tokenizer(data.prompt, truncation=True, max_length=512)
+        inputs['max_tokens'] = data.max_tokens
+        inputs['top_k'] = data.top_k
+        inputs['top_p'] = data.top_p
+        inputs['temperature'] = data.temperature
+        try:
+            uid = id(data)
+            engine.submit(uid, inputs)
+            output = await engine.wait(uid)
+            output = tokenizer.decode(output, skip_special_tokens=True)
+            if cache is not None:
+                cache.add(key, output)
+        except QueueFullError as e:
+            raise HTTPException(status_code=406, detail=e.args[0])
+
+    return {'text': output}
+
+
+@app.on_event("shutdown")
+async def shutdown(*_):
+    engine.shutdown()
+    server.should_exit = True
+    server.force_exit = True
+    await server.shutdown()
+
+
+def get_model_fn(model_name: str):
+    model_map = {
+        'opt-125m': opt_125M,
+        'opt-6.7b': opt_6B,
+        'opt-30b': opt_30B,
+        'opt-175b': opt_175B
+    }
+    return model_map[model_name]
+
+
+def print_args(args: argparse.Namespace):
+    print('\n==> Args:')
+    for k, v in args.__dict__.items():
+        print(f'{k} = {v}')
+
+
+FIXED_CACHE_KEYS = [
+    ('Question: What is the name of the largest continent on earth?\nAnswer: Asia\n\nQuestion: What is at the center of the solar system?\nAnswer:', 64),
+    ('A chat between a salesman and a student.\n\nSalesman: Hi boy, are you looking for a new phone?\nStudent: Yes, my phone is not functioning well.\nSalesman: What is your budget? \nStudent: I have received my scholarship so I am fine with any phone.\nSalesman: Great, then perhaps this latest flagship phone is just right for you.', 64),
+    ("English: I am happy today.\nChinese: 我今天很开心。\n\nEnglish: I am going to play basketball.\nChinese: 我一会去打篮球。\n\nEnglish: Let's celebrate our anniversary.\nChinese:", 64)
+]
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('model', choices=['opt-125m', 'opt-6.7b', 'opt-30b', 'opt-175b'])
+    parser.add_argument('--tp', type=int, default=1)
+    parser.add_argument('--master_host', default='localhost')
+    parser.add_argument('--master_port', type=int, default=19990)
+    parser.add_argument('--rpc_port', type=int, default=19980)
+    parser.add_argument('--max_batch_size', type=int, default=8)
+    parser.add_argument('--pipe_size', type=int, default=1)
+    parser.add_argument('--queue_size', type=int, default=0)
+    parser.add_argument('--http_host', default='0.0.0.0')
+    parser.add_argument('--http_port', type=int, default=7070)
+    parser.add_argument('--checkpoint', default=None)
+    parser.add_argument('--cache_size', type=int, default=0)
+    parser.add_argument('--cache_list_size', type=int, default=1)
+    args = parser.parse_args()
+    print_args(args)
+    model_kwargs = {}
+    if args.checkpoint is not None:
+        model_kwargs['checkpoint'] = args.checkpoint
+
+    logger = logging.getLogger(__name__)
+    tokenizer = GPT2Tokenizer.from_pretrained('facebook/opt-30b')
+    if args.cache_size > 0:
+        cache = ListCache(args.cache_size, args.cache_list_size, fixed_keys=FIXED_CACHE_KEYS)
+    else:
+        cache = None
+    engine = launch_engine(args.tp, 1, args.master_host, args.master_port, args.rpc_port, get_model_fn(args.model),
+                           batch_manager=BatchManagerForGeneration(max_batch_size=args.max_batch_size,
+                                                                   pad_token_id=tokenizer.pad_token_id),
+                           pipe_size=args.pipe_size,
+                           queue_size=args.queue_size,
+                           **model_kwargs)
+    config = uvicorn.Config(app, host=args.http_host, port=args.http_port)
+    server = uvicorn.Server(config=config)
+    server.run()

examples/tutorial/opt/inference/opt_server.py

+import logging
+import argparse
+import random
+from torch import Tensor
+from pydantic import BaseModel, Field
+from typing import Optional
+from energonai.model import opt_125M, opt_30B, opt_175B, opt_6B
+from transformers import GPT2Tokenizer
+from energonai import launch_engine, QueueFullError
+from sanic import Sanic
+from sanic.request import Request
+from sanic.response import json
+from sanic_ext import validate, openapi
+from batch import BatchManagerForGeneration
+from cache import ListCache, MissCacheError
+
+
+class GenerationTaskReq(BaseModel):
+    max_tokens: int = Field(gt=0, le=256, example=64)
+    prompt: str = Field(
+        min_length=1, example='Question: Where were the 2004 Olympics held?\nAnswer: Athens, Greece\n\nQuestion: What is the longest river on the earth?\nAnswer:')
+    top_k: Optional[int] = Field(default=None, gt=0, example=50)
+    top_p: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.5)
+    temperature: Optional[float] = Field(default=None, gt=0.0, lt=1.0, example=0.7)
+
+
+app = Sanic('opt')
+
+
+@app.post('/generation')
+@openapi.body(GenerationTaskReq)
+@validate(json=GenerationTaskReq)
+async def generate(request: Request, body: GenerationTaskReq):
+    logger.info(f'{request.ip}:{request.port} - "{request.method} {request.path}" - {body}')
+    key = (body.prompt, body.max_tokens)
+    try:
+        if cache is None:
+            raise MissCacheError()
+        outputs = cache.get(key)
+        output = random.choice(outputs)
+        logger.info('Cache hit')
+    except MissCacheError:
+        inputs = tokenizer(body.prompt, truncation=True, max_length=512)
+        inputs['max_tokens'] = body.max_tokens
+        inputs['top_k'] = body.top_k
+        inputs['top_p'] = body.top_p
+        inputs['temperature'] = body.temperature
+        try:
+            uid = id(body)
+            engine.submit(uid, inputs)
+            output = await engine.wait(uid)
+            assert isinstance(output, Tensor)
+            output = tokenizer.decode(output, skip_special_tokens=True)
+            if cache is not None:
+                cache.add(key, output)
+        except QueueFullError as e:
+            return json({'detail': e.args[0]}, status=406)
+
+    return json({'text': output})
+
+
+@app.after_server_stop
+def shutdown(*_):
+    engine.shutdown()
+
+
+def get_model_fn(model_name: str):
+    model_map = {
+        'opt-125m': opt_125M,
+        'opt-6.7b': opt_6B,
+        'opt-30b': opt_30B,
+        'opt-175b': opt_175B
+    }
+    return model_map[model_name]
+
+
+def print_args(args: argparse.Namespace):
+    print('\n==> Args:')
+    for k, v in args.__dict__.items():
+        print(f'{k} = {v}')
+
+
+FIXED_CACHE_KEYS = [
+    ('Question: What is the name of the largest continent on earth?\nAnswer: Asia\n\nQuestion: What is at the center of the solar system?\nAnswer:', 64),
+    ('A chat between a salesman and a student.\n\nSalesman: Hi boy, are you looking for a new phone?\nStudent: Yes, my phone is not functioning well.\nSalesman: What is your budget? \nStudent: I have received my scholarship so I am fine with any phone.\nSalesman: Great, then perhaps this latest flagship phone is just right for you.', 64),
+    ("English: I am happy today.\nChinese: 我今天很开心。\n\nEnglish: I am going to play basketball.\nChinese: 我一会去打篮球。\n\nEnglish: Let's celebrate our anniversary.\nChinese:", 64)
+]
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('model', choices=['opt-125m', 'opt-6.7b', 'opt-30b', 'opt-175b'])
+    parser.add_argument('--tp', type=int, default=1)
+    parser.add_argument('--master_host', default='localhost')
+    parser.add_argument('--master_port', type=int, default=19990)
+    parser.add_argument('--rpc_port', type=int, default=19980)
+    parser.add_argument('--max_batch_size', type=int, default=8)
+    parser.add_argument('--pipe_size', type=int, default=1)
+    parser.add_argument('--queue_size', type=int, default=0)
+    parser.add_argument('--http_host', default='0.0.0.0')
+    parser.add_argument('--http_port', type=int, default=7070)
+    parser.add_argument('--checkpoint', default=None)
+    parser.add_argument('--cache_size', type=int, default=0)
+    parser.add_argument('--cache_list_size', type=int, default=1)
+    args = parser.parse_args()
+    print_args(args)
+    model_kwargs = {}
+    if args.checkpoint is not None:
+        model_kwargs['checkpoint'] = args.checkpoint
+
+    logger = logging.getLogger(__name__)
+    tokenizer = GPT2Tokenizer.from_pretrained('facebook/opt-30b')
+    if args.cache_size > 0:
+        cache = ListCache(args.cache_size, args.cache_list_size, fixed_keys=FIXED_CACHE_KEYS)
+    else:
+        cache = None
+    engine = launch_engine(args.tp, 1, args.master_host, args.master_port, args.rpc_port, get_model_fn(args.model),
+                           batch_manager=BatchManagerForGeneration(max_batch_size=args.max_batch_size,
+                                                                   pad_token_id=tokenizer.pad_token_id),
+                           pipe_size=args.pipe_size,
+                           queue_size=args.queue_size,
+                           **model_kwargs)
+    app.run(args.http_host, args.http_port)

examples/tutorial/opt/inference/requirements.txt

+fastapi==0.85.1
+locust==2.11.0
+pydantic==1.10.2
+sanic==22.9.0
+sanic_ext==22.9.0
+torch>=1.10.0
+transformers==4.23.1
+uvicorn==0.19.0

examples/tutorial/opt/inference/script/process-opt-175b/README.md

+# Process OPT-175B weights
+
+You should download the pre-trained weights following the [doc](https://github.com/facebookresearch/metaseq/tree/main/projects/OPT) before reading this.
+
+First, install `metaseq` and `git clone https://github.com/facebookresearch/metaseq.git`.
+
+Then, `cd metaseq`.
+
+To consolidate checkpoints to eliminate FSDP:
+
+```shell
+bash metaseq/scripts/reshard_mp_launch_no_slurm.sh <directory_where_all_the_shards_are>/checkpoint_last <output_dir>/ 8 1
+```
+
+You will get 8 files in `<output_dir>`, and you should have the following checksums:
+```
+7e71cb65c4be784aa0b2889ac6039ee8  reshard-model_part-0-shard0.pt
+c8123da04f2c25a9026ea3224d5d5022  reshard-model_part-1-shard0.pt
+45e5d10896382e5bc4a7064fcafd2b1e  reshard-model_part-2-shard0.pt
+abb7296c4d2fc17420b84ca74fc3ce64  reshard-model_part-3-shard0.pt
+05dcc7ac6046f4d3f90b3d1068e6da15  reshard-model_part-4-shard0.pt
+d24dd334019060ce1ee7e625fcf6b4bd  reshard-model_part-5-shard0.pt
+fb1615ce0bbe89cc717f3e5079ee2655  reshard-model_part-6-shard0.pt
+2f3124432d2dbc6aebfca06be4b791c2  reshard-model_part-7-shard0.pt
+```
+
+Copy `flat-meta.json` to `<output_dir>`.
+
+Then cd to this dir, and we unflatten parameters.
+
+```shell
+bash unflat.sh <output_dir>/ <new_output_dir>/
+```
+
+Finally, you will get 8 files in `<new_output_dir>` with following checksums:
+```
+6169c59d014be95553c89ec01b8abb62  reshard-model_part-0.pt
+58868105da3d74a528a548fdb3a8cff6  reshard-model_part-1.pt
+69b255dc5a49d0eba9e4b60432cda90b  reshard-model_part-2.pt
+002c052461ff9ffb0cdac3d5906f41f2  reshard-model_part-3.pt
+6d57f72909320d511ffd5f1c668b2beb  reshard-model_part-4.pt
+93c8c4041cdc0c7907cc7afcf15cec2a  reshard-model_part-5.pt
+5d63b8750d827a1aa7c8ae5b02a3a2ca  reshard-model_part-6.pt
+f888bd41e009096804fe9a4b48c7ffe8  reshard-model_part-7.pt
+```
+

examples/tutorial/opt/inference/script/process-opt-175b/convert_ckpt.py

+import argparse
+import json
+import os
+import re
+from collections import defaultdict
+
+import numpy as np
+import torch
+
+
+def load_json(path: str):
+    with open(path) as f:
+        return json.load(f)
+
+
+def parse_shape_info(flat_dir: str):
+    data = load_json(os.path.join(flat_dir, 'shape.json'))
+    flat_info = defaultdict(lambda: defaultdict(list))
+    for k, shape in data.items():
+        matched = re.match(r'decoder.layers.\d+', k)
+        if matched is None:
+            flat_key = 'flat_param_0'
+        else:
+            flat_key = f'{matched[0]}.flat_param_0'
+        flat_info[flat_key]['names'].append(k)
+        flat_info[flat_key]['shapes'].append(shape)
+        flat_info[flat_key]['numels'].append(int(np.prod(shape)))
+    return flat_info
+
+
+def convert(flat_dir: str, output_dir: str, part: int):
+    flat_path = os.path.join(flat_dir, f'reshard-model_part-{part}-shard0.pt')
+    output_path = os.path.join(output_dir, f'reshard-model_part-{part}.pt')
+    flat_meta = load_json(os.path.join(flat_dir, 'flat-meta.json'))
+    flat_sd = torch.load(flat_path)
+    print(f'Loaded flat state dict from {flat_path}')
+    output_sd = {}
+    for flat_key, param_meta in flat_meta.items():
+        flat_param = flat_sd['model'][flat_key]
+        assert sum(param_meta['numels']) == flat_param.numel(
+        ), f'flat {flat_key} {flat_param.numel()} vs {sum(param_meta["numels"])}'
+        for name, shape, param in zip(param_meta['names'], param_meta['shapes'], flat_param.split(param_meta['numels'])):
+            output_sd[name] = param.view(shape)
+
+    torch.save(output_sd, output_path)
+    print(f'Saved unflat state dict to {output_path}')
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('flat_dir')
+    parser.add_argument('output_dir')
+    parser.add_argument('part', type=int)
+    args = parser.parse_args()
+    convert(args.flat_dir, args.output_dir, args.part)

examples/tutorial/opt/inference/script/process-opt-175b/flat-meta.json

+{"flat_param_0": {"names": ["decoder.embed_tokens.weight", "decoder.embed_positions.weight", "decoder.layer_norm.weight", "decoder.layer_norm.bias"], "shapes": [[6284, 12288], [2050, 12288], [12288], [12288]], "numels": [77217792, 25190400, 12288, 12288]}, "decoder.layers.0.flat_param_0": {"names": ["decoder.layers.0.self_attn.qkv_proj.weight", "decoder.layers.0.self_attn.qkv_proj.bias", "decoder.layers.0.self_attn.out_proj.weight", "decoder.layers.0.self_attn.out_proj.bias", "decoder.layers.0.self_attn_layer_norm.weight", "decoder.layers.0.self_attn_layer_norm.bias", "decoder.layers.0.fc1.weight", "decoder.layers.0.fc1.bias", "decoder.layers.0.fc2.weight", "decoder.layers.0.fc2.bias", "decoder.layers.0.final_layer_norm.weight", "decoder.layers.0.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.1.flat_param_0": {"names": ["decoder.layers.1.self_attn.qkv_proj.weight", "decoder.layers.1.self_attn.qkv_proj.bias", "decoder.layers.1.self_attn.out_proj.weight", "decoder.layers.1.self_attn.out_proj.bias", "decoder.layers.1.self_attn_layer_norm.weight", "decoder.layers.1.self_attn_layer_norm.bias", "decoder.layers.1.fc1.weight", "decoder.layers.1.fc1.bias", "decoder.layers.1.fc2.weight", "decoder.layers.1.fc2.bias", "decoder.layers.1.final_layer_norm.weight", "decoder.layers.1.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.2.flat_param_0": {"names": ["decoder.layers.2.self_attn.qkv_proj.weight", "decoder.layers.2.self_attn.qkv_proj.bias", "decoder.layers.2.self_attn.out_proj.weight", "decoder.layers.2.self_attn.out_proj.bias", "decoder.layers.2.self_attn_layer_norm.weight", "decoder.layers.2.self_attn_layer_norm.bias", "decoder.layers.2.fc1.weight", "decoder.layers.2.fc1.bias", "decoder.layers.2.fc2.weight", "decoder.layers.2.fc2.bias", "decoder.layers.2.final_layer_norm.weight", "decoder.layers.2.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.3.flat_param_0": {"names": ["decoder.layers.3.self_attn.qkv_proj.weight", "decoder.layers.3.self_attn.qkv_proj.bias", "decoder.layers.3.self_attn.out_proj.weight", "decoder.layers.3.self_attn.out_proj.bias", "decoder.layers.3.self_attn_layer_norm.weight", "decoder.layers.3.self_attn_layer_norm.bias", "decoder.layers.3.fc1.weight", "decoder.layers.3.fc1.bias", "decoder.layers.3.fc2.weight", "decoder.layers.3.fc2.bias", "decoder.layers.3.final_layer_norm.weight", "decoder.layers.3.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.4.flat_param_0": {"names": ["decoder.layers.4.self_attn.qkv_proj.weight", "decoder.layers.4.self_attn.qkv_proj.bias", "decoder.layers.4.self_attn.out_proj.weight", "decoder.layers.4.self_attn.out_proj.bias", "decoder.layers.4.self_attn_layer_norm.weight", "decoder.layers.4.self_attn_layer_norm.bias", "decoder.layers.4.fc1.weight", "decoder.layers.4.fc1.bias", "decoder.layers.4.fc2.weight", "decoder.layers.4.fc2.bias", "decoder.layers.4.final_layer_norm.weight", "decoder.layers.4.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.5.flat_param_0": {"names": ["decoder.layers.5.self_attn.qkv_proj.weight", "decoder.layers.5.self_attn.qkv_proj.bias", "decoder.layers.5.self_attn.out_proj.weight", "decoder.layers.5.self_attn.out_proj.bias", "decoder.layers.5.self_attn_layer_norm.weight", "decoder.layers.5.self_attn_layer_norm.bias", "decoder.layers.5.fc1.weight", "decoder.layers.5.fc1.bias", "decoder.layers.5.fc2.weight", "decoder.layers.5.fc2.bias", "decoder.layers.5.final_layer_norm.weight", "decoder.layers.5.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.6.flat_param_0": {"names": ["decoder.layers.6.self_attn.qkv_proj.weight", "decoder.layers.6.self_attn.qkv_proj.bias", "decoder.layers.6.self_attn.out_proj.weight", "decoder.layers.6.self_attn.out_proj.bias", "decoder.layers.6.self_attn_layer_norm.weight", "decoder.layers.6.self_attn_layer_norm.bias", "decoder.layers.6.fc1.weight", "decoder.layers.6.fc1.bias", "decoder.layers.6.fc2.weight", "decoder.layers.6.fc2.bias", "decoder.layers.6.final_layer_norm.weight", "decoder.layers.6.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.7.flat_param_0": {"names": ["decoder.layers.7.self_attn.qkv_proj.weight", "decoder.layers.7.self_attn.qkv_proj.bias", "decoder.layers.7.self_attn.out_proj.weight", "decoder.layers.7.self_attn.out_proj.bias", "decoder.layers.7.self_attn_layer_norm.weight", "decoder.layers.7.self_attn_layer_norm.bias", "decoder.layers.7.fc1.weight", "decoder.layers.7.fc1.bias", "decoder.layers.7.fc2.weight", "decoder.layers.7.fc2.bias", "decoder.layers.7.final_layer_norm.weight", "decoder.layers.7.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.8.flat_param_0": {"names": ["decoder.layers.8.self_attn.qkv_proj.weight", "decoder.layers.8.self_attn.qkv_proj.bias", "decoder.layers.8.self_attn.out_proj.weight", "decoder.layers.8.self_attn.out_proj.bias", "decoder.layers.8.self_attn_layer_norm.weight", "decoder.layers.8.self_attn_layer_norm.bias", "decoder.layers.8.fc1.weight", "decoder.layers.8.fc1.bias", "decoder.layers.8.fc2.weight", "decoder.layers.8.fc2.bias", "decoder.layers.8.final_layer_norm.weight", "decoder.layers.8.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.9.flat_param_0": {"names": ["decoder.layers.9.self_attn.qkv_proj.weight", "decoder.layers.9.self_attn.qkv_proj.bias", "decoder.layers.9.self_attn.out_proj.weight", "decoder.layers.9.self_attn.out_proj.bias", "decoder.layers.9.self_attn_layer_norm.weight", "decoder.layers.9.self_attn_layer_norm.bias", "decoder.layers.9.fc1.weight", "decoder.layers.9.fc1.bias", "decoder.layers.9.fc2.weight", "decoder.layers.9.fc2.bias", "decoder.layers.9.final_layer_norm.weight", "decoder.layers.9.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.10.flat_param_0": {"names": ["decoder.layers.10.self_attn.qkv_proj.weight", "decoder.layers.10.self_attn.qkv_proj.bias", "decoder.layers.10.self_attn.out_proj.weight", "decoder.layers.10.self_attn.out_proj.bias", "decoder.layers.10.self_attn_layer_norm.weight", "decoder.layers.10.self_attn_layer_norm.bias", "decoder.layers.10.fc1.weight", "decoder.layers.10.fc1.bias", "decoder.layers.10.fc2.weight", "decoder.layers.10.fc2.bias", "decoder.layers.10.final_layer_norm.weight", "decoder.layers.10.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.11.flat_param_0": {"names": ["decoder.layers.11.self_attn.qkv_proj.weight", "decoder.layers.11.self_attn.qkv_proj.bias", "decoder.layers.11.self_attn.out_proj.weight", "decoder.layers.11.self_attn.out_proj.bias", "decoder.layers.11.self_attn_layer_norm.weight", "decoder.layers.11.self_attn_layer_norm.bias", "decoder.layers.11.fc1.weight", "decoder.layers.11.fc1.bias", "decoder.layers.11.fc2.weight", "decoder.layers.11.fc2.bias", "decoder.layers.11.final_layer_norm.weight", "decoder.layers.11.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.12.flat_param_0": {"names": ["decoder.layers.12.self_attn.qkv_proj.weight", "decoder.layers.12.self_attn.qkv_proj.bias", "decoder.layers.12.self_attn.out_proj.weight", "decoder.layers.12.self_attn.out_proj.bias", "decoder.layers.12.self_attn_layer_norm.weight", "decoder.layers.12.self_attn_layer_norm.bias", "decoder.layers.12.fc1.weight", "decoder.layers.12.fc1.bias", "decoder.layers.12.fc2.weight", "decoder.layers.12.fc2.bias", "decoder.layers.12.final_layer_norm.weight", "decoder.layers.12.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.13.flat_param_0": {"names": ["decoder.layers.13.self_attn.qkv_proj.weight", "decoder.layers.13.self_attn.qkv_proj.bias", "decoder.layers.13.self_attn.out_proj.weight", "decoder.layers.13.self_attn.out_proj.bias", "decoder.layers.13.self_attn_layer_norm.weight", "decoder.layers.13.self_attn_layer_norm.bias", "decoder.layers.13.fc1.weight", "decoder.layers.13.fc1.bias", "decoder.layers.13.fc2.weight", "decoder.layers.13.fc2.bias", "decoder.layers.13.final_layer_norm.weight", "decoder.layers.13.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.14.flat_param_0": {"names": ["decoder.layers.14.self_attn.qkv_proj.weight", "decoder.layers.14.self_attn.qkv_proj.bias", "decoder.layers.14.self_attn.out_proj.weight", "decoder.layers.14.self_attn.out_proj.bias", "decoder.layers.14.self_attn_layer_norm.weight", "decoder.layers.14.self_attn_layer_norm.bias", "decoder.layers.14.fc1.weight", "decoder.layers.14.fc1.bias", "decoder.layers.14.fc2.weight", "decoder.layers.14.fc2.bias", "decoder.layers.14.final_layer_norm.weight", "decoder.layers.14.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.15.flat_param_0": {"names": ["decoder.layers.15.self_attn.qkv_proj.weight", "decoder.layers.15.self_attn.qkv_proj.bias", "decoder.layers.15.self_attn.out_proj.weight", "decoder.layers.15.self_attn.out_proj.bias", "decoder.layers.15.self_attn_layer_norm.weight", "decoder.layers.15.self_attn_layer_norm.bias", "decoder.layers.15.fc1.weight", "decoder.layers.15.fc1.bias", "decoder.layers.15.fc2.weight", "decoder.layers.15.fc2.bias", "decoder.layers.15.final_layer_norm.weight", "decoder.layers.15.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.16.flat_param_0": {"names": ["decoder.layers.16.self_attn.qkv_proj.weight", "decoder.layers.16.self_attn.qkv_proj.bias", "decoder.layers.16.self_attn.out_proj.weight", "decoder.layers.16.self_attn.out_proj.bias", "decoder.layers.16.self_attn_layer_norm.weight", "decoder.layers.16.self_attn_layer_norm.bias", "decoder.layers.16.fc1.weight", "decoder.layers.16.fc1.bias", "decoder.layers.16.fc2.weight", "decoder.layers.16.fc2.bias", "decoder.layers.16.final_layer_norm.weight", "decoder.layers.16.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.17.flat_param_0": {"names": ["decoder.layers.17.self_attn.qkv_proj.weight", "decoder.layers.17.self_attn.qkv_proj.bias", "decoder.layers.17.self_attn.out_proj.weight", "decoder.layers.17.self_attn.out_proj.bias", "decoder.layers.17.self_attn_layer_norm.weight", "decoder.layers.17.self_attn_layer_norm.bias", "decoder.layers.17.fc1.weight", "decoder.layers.17.fc1.bias", "decoder.layers.17.fc2.weight", "decoder.layers.17.fc2.bias", "decoder.layers.17.final_layer_norm.weight", "decoder.layers.17.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.18.flat_param_0": {"names": ["decoder.layers.18.self_attn.qkv_proj.weight", "decoder.layers.18.self_attn.qkv_proj.bias", "decoder.layers.18.self_attn.out_proj.weight", "decoder.layers.18.self_attn.out_proj.bias", "decoder.layers.18.self_attn_layer_norm.weight", "decoder.layers.18.self_attn_layer_norm.bias", "decoder.layers.18.fc1.weight", "decoder.layers.18.fc1.bias", "decoder.layers.18.fc2.weight", "decoder.layers.18.fc2.bias", "decoder.layers.18.final_layer_norm.weight", "decoder.layers.18.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.19.flat_param_0": {"names": ["decoder.layers.19.self_attn.qkv_proj.weight", "decoder.layers.19.self_attn.qkv_proj.bias", "decoder.layers.19.self_attn.out_proj.weight", "decoder.layers.19.self_attn.out_proj.bias", "decoder.layers.19.self_attn_layer_norm.weight", "decoder.layers.19.self_attn_layer_norm.bias", "decoder.layers.19.fc1.weight", "decoder.layers.19.fc1.bias", "decoder.layers.19.fc2.weight", "decoder.layers.19.fc2.bias", "decoder.layers.19.final_layer_norm.weight", "decoder.layers.19.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.20.flat_param_0": {"names": ["decoder.layers.20.self_attn.qkv_proj.weight", "decoder.layers.20.self_attn.qkv_proj.bias", "decoder.layers.20.self_attn.out_proj.weight", "decoder.layers.20.self_attn.out_proj.bias", "decoder.layers.20.self_attn_layer_norm.weight", "decoder.layers.20.self_attn_layer_norm.bias", "decoder.layers.20.fc1.weight", "decoder.layers.20.fc1.bias", "decoder.layers.20.fc2.weight", "decoder.layers.20.fc2.bias", "decoder.layers.20.final_layer_norm.weight", "decoder.layers.20.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.21.flat_param_0": {"names": ["decoder.layers.21.self_attn.qkv_proj.weight", "decoder.layers.21.self_attn.qkv_proj.bias", "decoder.layers.21.self_attn.out_proj.weight", "decoder.layers.21.self_attn.out_proj.bias", "decoder.layers.21.self_attn_layer_norm.weight", "decoder.layers.21.self_attn_layer_norm.bias", "decoder.layers.21.fc1.weight", "decoder.layers.21.fc1.bias", "decoder.layers.21.fc2.weight", "decoder.layers.21.fc2.bias", "decoder.layers.21.final_layer_norm.weight", "decoder.layers.21.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.22.flat_param_0": {"names": ["decoder.layers.22.self_attn.qkv_proj.weight", "decoder.layers.22.self_attn.qkv_proj.bias", "decoder.layers.22.self_attn.out_proj.weight", "decoder.layers.22.self_attn.out_proj.bias", "decoder.layers.22.self_attn_layer_norm.weight", "decoder.layers.22.self_attn_layer_norm.bias", "decoder.layers.22.fc1.weight", "decoder.layers.22.fc1.bias", "decoder.layers.22.fc2.weight", "decoder.layers.22.fc2.bias", "decoder.layers.22.final_layer_norm.weight", "decoder.layers.22.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.23.flat_param_0": {"names": ["decoder.layers.23.self_attn.qkv_proj.weight", "decoder.layers.23.self_attn.qkv_proj.bias", "decoder.layers.23.self_attn.out_proj.weight", "decoder.layers.23.self_attn.out_proj.bias", "decoder.layers.23.self_attn_layer_norm.weight", "decoder.layers.23.self_attn_layer_norm.bias", "decoder.layers.23.fc1.weight", "decoder.layers.23.fc1.bias", "decoder.layers.23.fc2.weight", "decoder.layers.23.fc2.bias", "decoder.layers.23.final_layer_norm.weight", "decoder.layers.23.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.24.flat_param_0": {"names": ["decoder.layers.24.self_attn.qkv_proj.weight", "decoder.layers.24.self_attn.qkv_proj.bias", "decoder.layers.24.self_attn.out_proj.weight", "decoder.layers.24.self_attn.out_proj.bias", "decoder.layers.24.self_attn_layer_norm.weight", "decoder.layers.24.self_attn_layer_norm.bias", "decoder.layers.24.fc1.weight", "decoder.layers.24.fc1.bias", "decoder.layers.24.fc2.weight", "decoder.layers.24.fc2.bias", "decoder.layers.24.final_layer_norm.weight", "decoder.layers.24.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.25.flat_param_0": {"names": ["decoder.layers.25.self_attn.qkv_proj.weight", "decoder.layers.25.self_attn.qkv_proj.bias", "decoder.layers.25.self_attn.out_proj.weight", "decoder.layers.25.self_attn.out_proj.bias", "decoder.layers.25.self_attn_layer_norm.weight", "decoder.layers.25.self_attn_layer_norm.bias", "decoder.layers.25.fc1.weight", "decoder.layers.25.fc1.bias", "decoder.layers.25.fc2.weight", "decoder.layers.25.fc2.bias", "decoder.layers.25.final_layer_norm.weight", "decoder.layers.25.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.26.flat_param_0": {"names": ["decoder.layers.26.self_attn.qkv_proj.weight", "decoder.layers.26.self_attn.qkv_proj.bias", "decoder.layers.26.self_attn.out_proj.weight", "decoder.layers.26.self_attn.out_proj.bias", "decoder.layers.26.self_attn_layer_norm.weight", "decoder.layers.26.self_attn_layer_norm.bias", "decoder.layers.26.fc1.weight", "decoder.layers.26.fc1.bias", "decoder.layers.26.fc2.weight", "decoder.layers.26.fc2.bias", "decoder.layers.26.final_layer_norm.weight", "decoder.layers.26.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.27.flat_param_0": {"names": ["decoder.layers.27.self_attn.qkv_proj.weight", "decoder.layers.27.self_attn.qkv_proj.bias", "decoder.layers.27.self_attn.out_proj.weight", "decoder.layers.27.self_attn.out_proj.bias", "decoder.layers.27.self_attn_layer_norm.weight", "decoder.layers.27.self_attn_layer_norm.bias", "decoder.layers.27.fc1.weight", "decoder.layers.27.fc1.bias", "decoder.layers.27.fc2.weight", "decoder.layers.27.fc2.bias", "decoder.layers.27.final_layer_norm.weight", "decoder.layers.27.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.28.flat_param_0": {"names": ["decoder.layers.28.self_attn.qkv_proj.weight", "decoder.layers.28.self_attn.qkv_proj.bias", "decoder.layers.28.self_attn.out_proj.weight", "decoder.layers.28.self_attn.out_proj.bias", "decoder.layers.28.self_attn_layer_norm.weight", "decoder.layers.28.self_attn_layer_norm.bias", "decoder.layers.28.fc1.weight", "decoder.layers.28.fc1.bias", "decoder.layers.28.fc2.weight", "decoder.layers.28.fc2.bias", "decoder.layers.28.final_layer_norm.weight", "decoder.layers.28.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.29.flat_param_0": {"names": ["decoder.layers.29.self_attn.qkv_proj.weight", "decoder.layers.29.self_attn.qkv_proj.bias", "decoder.layers.29.self_attn.out_proj.weight", "decoder.layers.29.self_attn.out_proj.bias", "decoder.layers.29.self_attn_layer_norm.weight", "decoder.layers.29.self_attn_layer_norm.bias", "decoder.layers.29.fc1.weight", "decoder.layers.29.fc1.bias", "decoder.layers.29.fc2.weight", "decoder.layers.29.fc2.bias", "decoder.layers.29.final_layer_norm.weight", "decoder.layers.29.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.30.flat_param_0": {"names": ["decoder.layers.30.self_attn.qkv_proj.weight", "decoder.layers.30.self_attn.qkv_proj.bias", "decoder.layers.30.self_attn.out_proj.weight", "decoder.layers.30.self_attn.out_proj.bias", "decoder.layers.30.self_attn_layer_norm.weight", "decoder.layers.30.self_attn_layer_norm.bias", "decoder.layers.30.fc1.weight", "decoder.layers.30.fc1.bias", "decoder.layers.30.fc2.weight", "decoder.layers.30.fc2.bias", "decoder.layers.30.final_layer_norm.weight", "decoder.layers.30.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.31.flat_param_0": {"names": ["decoder.layers.31.self_attn.qkv_proj.weight", "decoder.layers.31.self_attn.qkv_proj.bias", "decoder.layers.31.self_attn.out_proj.weight", "decoder.layers.31.self_attn.out_proj.bias", "decoder.layers.31.self_attn_layer_norm.weight", "decoder.layers.31.self_attn_layer_norm.bias", "decoder.layers.31.fc1.weight", "decoder.layers.31.fc1.bias", "decoder.layers.31.fc2.weight", "decoder.layers.31.fc2.bias", "decoder.layers.31.final_layer_norm.weight", "decoder.layers.31.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.32.flat_param_0": {"names": ["decoder.layers.32.self_attn.qkv_proj.weight", "decoder.layers.32.self_attn.qkv_proj.bias", "decoder.layers.32.self_attn.out_proj.weight", "decoder.layers.32.self_attn.out_proj.bias", "decoder.layers.32.self_attn_layer_norm.weight", "decoder.layers.32.self_attn_layer_norm.bias", "decoder.layers.32.fc1.weight", "decoder.layers.32.fc1.bias", "decoder.layers.32.fc2.weight", "decoder.layers.32.fc2.bias", "decoder.layers.32.final_layer_norm.weight", "decoder.layers.32.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.33.flat_param_0": {"names": ["decoder.layers.33.self_attn.qkv_proj.weight", "decoder.layers.33.self_attn.qkv_proj.bias", "decoder.layers.33.self_attn.out_proj.weight", "decoder.layers.33.self_attn.out_proj.bias", "decoder.layers.33.self_attn_layer_norm.weight", "decoder.layers.33.self_attn_layer_norm.bias", "decoder.layers.33.fc1.weight", "decoder.layers.33.fc1.bias", "decoder.layers.33.fc2.weight", "decoder.layers.33.fc2.bias", "decoder.layers.33.final_layer_norm.weight", "decoder.layers.33.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.34.flat_param_0": {"names": ["decoder.layers.34.self_attn.qkv_proj.weight", "decoder.layers.34.self_attn.qkv_proj.bias", "decoder.layers.34.self_attn.out_proj.weight", "decoder.layers.34.self_attn.out_proj.bias", "decoder.layers.34.self_attn_layer_norm.weight", "decoder.layers.34.self_attn_layer_norm.bias", "decoder.layers.34.fc1.weight", "decoder.layers.34.fc1.bias", "decoder.layers.34.fc2.weight", "decoder.layers.34.fc2.bias", "decoder.layers.34.final_layer_norm.weight", "decoder.layers.34.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.35.flat_param_0": {"names": ["decoder.layers.35.self_attn.qkv_proj.weight", "decoder.layers.35.self_attn.qkv_proj.bias", "decoder.layers.35.self_attn.out_proj.weight", "decoder.layers.35.self_attn.out_proj.bias", "decoder.layers.35.self_attn_layer_norm.weight", "decoder.layers.35.self_attn_layer_norm.bias", "decoder.layers.35.fc1.weight", "decoder.layers.35.fc1.bias", "decoder.layers.35.fc2.weight", "decoder.layers.35.fc2.bias", "decoder.layers.35.final_layer_norm.weight", "decoder.layers.35.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.36.flat_param_0": {"names": ["decoder.layers.36.self_attn.qkv_proj.weight", "decoder.layers.36.self_attn.qkv_proj.bias", "decoder.layers.36.self_attn.out_proj.weight", "decoder.layers.36.self_attn.out_proj.bias", "decoder.layers.36.self_attn_layer_norm.weight", "decoder.layers.36.self_attn_layer_norm.bias", "decoder.layers.36.fc1.weight", "decoder.layers.36.fc1.bias", "decoder.layers.36.fc2.weight", "decoder.layers.36.fc2.bias", "decoder.layers.36.final_layer_norm.weight", "decoder.layers.36.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.37.flat_param_0": {"names": ["decoder.layers.37.self_attn.qkv_proj.weight", "decoder.layers.37.self_attn.qkv_proj.bias", "decoder.layers.37.self_attn.out_proj.weight", "decoder.layers.37.self_attn.out_proj.bias", "decoder.layers.37.self_attn_layer_norm.weight", "decoder.layers.37.self_attn_layer_norm.bias", "decoder.layers.37.fc1.weight", "decoder.layers.37.fc1.bias", "decoder.layers.37.fc2.weight", "decoder.layers.37.fc2.bias", "decoder.layers.37.final_layer_norm.weight", "decoder.layers.37.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.38.flat_param_0": {"names": ["decoder.layers.38.self_attn.qkv_proj.weight", "decoder.layers.38.self_attn.qkv_proj.bias", "decoder.layers.38.self_attn.out_proj.weight", "decoder.layers.38.self_attn.out_proj.bias", "decoder.layers.38.self_attn_layer_norm.weight", "decoder.layers.38.self_attn_layer_norm.bias", "decoder.layers.38.fc1.weight", "decoder.layers.38.fc1.bias", "decoder.layers.38.fc2.weight", "decoder.layers.38.fc2.bias", "decoder.layers.38.final_layer_norm.weight", "decoder.layers.38.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.39.flat_param_0": {"names": ["decoder.layers.39.self_attn.qkv_proj.weight", "decoder.layers.39.self_attn.qkv_proj.bias", "decoder.layers.39.self_attn.out_proj.weight", "decoder.layers.39.self_attn.out_proj.bias", "decoder.layers.39.self_attn_layer_norm.weight", "decoder.layers.39.self_attn_layer_norm.bias", "decoder.layers.39.fc1.weight", "decoder.layers.39.fc1.bias", "decoder.layers.39.fc2.weight", "decoder.layers.39.fc2.bias", "decoder.layers.39.final_layer_norm.weight", "decoder.layers.39.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.40.flat_param_0": {"names": ["decoder.layers.40.self_attn.qkv_proj.weight", "decoder.layers.40.self_attn.qkv_proj.bias", "decoder.layers.40.self_attn.out_proj.weight", "decoder.layers.40.self_attn.out_proj.bias", "decoder.layers.40.self_attn_layer_norm.weight", "decoder.layers.40.self_attn_layer_norm.bias", "decoder.layers.40.fc1.weight", "decoder.layers.40.fc1.bias", "decoder.layers.40.fc2.weight", "decoder.layers.40.fc2.bias", "decoder.layers.40.final_layer_norm.weight", "decoder.layers.40.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.41.flat_param_0": {"names": ["decoder.layers.41.self_attn.qkv_proj.weight", "decoder.layers.41.self_attn.qkv_proj.bias", "decoder.layers.41.self_attn.out_proj.weight", "decoder.layers.41.self_attn.out_proj.bias", "decoder.layers.41.self_attn_layer_norm.weight", "decoder.layers.41.self_attn_layer_norm.bias", "decoder.layers.41.fc1.weight", "decoder.layers.41.fc1.bias", "decoder.layers.41.fc2.weight", "decoder.layers.41.fc2.bias", "decoder.layers.41.final_layer_norm.weight", "decoder.layers.41.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.42.flat_param_0": {"names": ["decoder.layers.42.self_attn.qkv_proj.weight", "decoder.layers.42.self_attn.qkv_proj.bias", "decoder.layers.42.self_attn.out_proj.weight", "decoder.layers.42.self_attn.out_proj.bias", "decoder.layers.42.self_attn_layer_norm.weight", "decoder.layers.42.self_attn_layer_norm.bias", "decoder.layers.42.fc1.weight", "decoder.layers.42.fc1.bias", "decoder.layers.42.fc2.weight", "decoder.layers.42.fc2.bias", "decoder.layers.42.final_layer_norm.weight", "decoder.layers.42.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.43.flat_param_0": {"names": ["decoder.layers.43.self_attn.qkv_proj.weight", "decoder.layers.43.self_attn.qkv_proj.bias", "decoder.layers.43.self_attn.out_proj.weight", "decoder.layers.43.self_attn.out_proj.bias", "decoder.layers.43.self_attn_layer_norm.weight", "decoder.layers.43.self_attn_layer_norm.bias", "decoder.layers.43.fc1.weight", "decoder.layers.43.fc1.bias", "decoder.layers.43.fc2.weight", "decoder.layers.43.fc2.bias", "decoder.layers.43.final_layer_norm.weight", "decoder.layers.43.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.44.flat_param_0": {"names": ["decoder.layers.44.self_attn.qkv_proj.weight", "decoder.layers.44.self_attn.qkv_proj.bias", "decoder.layers.44.self_attn.out_proj.weight", "decoder.layers.44.self_attn.out_proj.bias", "decoder.layers.44.self_attn_layer_norm.weight", "decoder.layers.44.self_attn_layer_norm.bias", "decoder.layers.44.fc1.weight", "decoder.layers.44.fc1.bias", "decoder.layers.44.fc2.weight", "decoder.layers.44.fc2.bias", "decoder.layers.44.final_layer_norm.weight", "decoder.layers.44.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.45.flat_param_0": {"names": ["decoder.layers.45.self_attn.qkv_proj.weight", "decoder.layers.45.self_attn.qkv_proj.bias", "decoder.layers.45.self_attn.out_proj.weight", "decoder.layers.45.self_attn.out_proj.bias", "decoder.layers.45.self_attn_layer_norm.weight", "decoder.layers.45.self_attn_layer_norm.bias", "decoder.layers.45.fc1.weight", "decoder.layers.45.fc1.bias", "decoder.layers.45.fc2.weight", "decoder.layers.45.fc2.bias", "decoder.layers.45.final_layer_norm.weight", "decoder.layers.45.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.46.flat_param_0": {"names": ["decoder.layers.46.self_attn.qkv_proj.weight", "decoder.layers.46.self_attn.qkv_proj.bias", "decoder.layers.46.self_attn.out_proj.weight", "decoder.layers.46.self_attn.out_proj.bias", "decoder.layers.46.self_attn_layer_norm.weight", "decoder.layers.46.self_attn_layer_norm.bias", "decoder.layers.46.fc1.weight", "decoder.layers.46.fc1.bias", "decoder.layers.46.fc2.weight", "decoder.layers.46.fc2.bias", "decoder.layers.46.final_layer_norm.weight", "decoder.layers.46.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.47.flat_param_0": {"names": ["decoder.layers.47.self_attn.qkv_proj.weight", "decoder.layers.47.self_attn.qkv_proj.bias", "decoder.layers.47.self_attn.out_proj.weight", "decoder.layers.47.self_attn.out_proj.bias", "decoder.layers.47.self_attn_layer_norm.weight", "decoder.layers.47.self_attn_layer_norm.bias", "decoder.layers.47.fc1.weight", "decoder.layers.47.fc1.bias", "decoder.layers.47.fc2.weight", "decoder.layers.47.fc2.bias", "decoder.layers.47.final_layer_norm.weight", "decoder.layers.47.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.48.flat_param_0": {"names": ["decoder.layers.48.self_attn.qkv_proj.weight", "decoder.layers.48.self_attn.qkv_proj.bias", "decoder.layers.48.self_attn.out_proj.weight", "decoder.layers.48.self_attn.out_proj.bias", "decoder.layers.48.self_attn_layer_norm.weight", "decoder.layers.48.self_attn_layer_norm.bias", "decoder.layers.48.fc1.weight", "decoder.layers.48.fc1.bias", "decoder.layers.48.fc2.weight", "decoder.layers.48.fc2.bias", "decoder.layers.48.final_layer_norm.weight", "decoder.layers.48.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.49.flat_param_0": {"names": ["decoder.layers.49.self_attn.qkv_proj.weight", "decoder.layers.49.self_attn.qkv_proj.bias", "decoder.layers.49.self_attn.out_proj.weight", "decoder.layers.49.self_attn.out_proj.bias", "decoder.layers.49.self_attn_layer_norm.weight", "decoder.layers.49.self_attn_layer_norm.bias", "decoder.layers.49.fc1.weight", "decoder.layers.49.fc1.bias", "decoder.layers.49.fc2.weight", "decoder.layers.49.fc2.bias", "decoder.layers.49.final_layer_norm.weight", "decoder.layers.49.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.50.flat_param_0": {"names": ["decoder.layers.50.self_attn.qkv_proj.weight", "decoder.layers.50.self_attn.qkv_proj.bias", "decoder.layers.50.self_attn.out_proj.weight", "decoder.layers.50.self_attn.out_proj.bias", "decoder.layers.50.self_attn_layer_norm.weight", "decoder.layers.50.self_attn_layer_norm.bias", "decoder.layers.50.fc1.weight", "decoder.layers.50.fc1.bias", "decoder.layers.50.fc2.weight", "decoder.layers.50.fc2.bias", "decoder.layers.50.final_layer_norm.weight", "decoder.layers.50.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.51.flat_param_0": {"names": ["decoder.layers.51.self_attn.qkv_proj.weight", "decoder.layers.51.self_attn.qkv_proj.bias", "decoder.layers.51.self_attn.out_proj.weight", "decoder.layers.51.self_attn.out_proj.bias", "decoder.layers.51.self_attn_layer_norm.weight", "decoder.layers.51.self_attn_layer_norm.bias", "decoder.layers.51.fc1.weight", "decoder.layers.51.fc1.bias", "decoder.layers.51.fc2.weight", "decoder.layers.51.fc2.bias", "decoder.layers.51.final_layer_norm.weight", "decoder.layers.51.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.52.flat_param_0": {"names": ["decoder.layers.52.self_attn.qkv_proj.weight", "decoder.layers.52.self_attn.qkv_proj.bias", "decoder.layers.52.self_attn.out_proj.weight", "decoder.layers.52.self_attn.out_proj.bias", "decoder.layers.52.self_attn_layer_norm.weight", "decoder.layers.52.self_attn_layer_norm.bias", "decoder.layers.52.fc1.weight", "decoder.layers.52.fc1.bias", "decoder.layers.52.fc2.weight", "decoder.layers.52.fc2.bias", "decoder.layers.52.final_layer_norm.weight", "decoder.layers.52.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.53.flat_param_0": {"names": ["decoder.layers.53.self_attn.qkv_proj.weight", "decoder.layers.53.self_attn.qkv_proj.bias", "decoder.layers.53.self_attn.out_proj.weight", "decoder.layers.53.self_attn.out_proj.bias", "decoder.layers.53.self_attn_layer_norm.weight", "decoder.layers.53.self_attn_layer_norm.bias", "decoder.layers.53.fc1.weight", "decoder.layers.53.fc1.bias", "decoder.layers.53.fc2.weight", "decoder.layers.53.fc2.bias", "decoder.layers.53.final_layer_norm.weight", "decoder.layers.53.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.54.flat_param_0": {"names": ["decoder.layers.54.self_attn.qkv_proj.weight", "decoder.layers.54.self_attn.qkv_proj.bias", "decoder.layers.54.self_attn.out_proj.weight", "decoder.layers.54.self_attn.out_proj.bias", "decoder.layers.54.self_attn_layer_norm.weight", "decoder.layers.54.self_attn_layer_norm.bias", "decoder.layers.54.fc1.weight", "decoder.layers.54.fc1.bias", "decoder.layers.54.fc2.weight", "decoder.layers.54.fc2.bias", "decoder.layers.54.final_layer_norm.weight", "decoder.layers.54.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.55.flat_param_0": {"names": ["decoder.layers.55.self_attn.qkv_proj.weight", "decoder.layers.55.self_attn.qkv_proj.bias", "decoder.layers.55.self_attn.out_proj.weight", "decoder.layers.55.self_attn.out_proj.bias", "decoder.layers.55.self_attn_layer_norm.weight", "decoder.layers.55.self_attn_layer_norm.bias", "decoder.layers.55.fc1.weight", "decoder.layers.55.fc1.bias", "decoder.layers.55.fc2.weight", "decoder.layers.55.fc2.bias", "decoder.layers.55.final_layer_norm.weight", "decoder.layers.55.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.56.flat_param_0": {"names": ["decoder.layers.56.self_attn.qkv_proj.weight", "decoder.layers.56.self_attn.qkv_proj.bias", "decoder.layers.56.self_attn.out_proj.weight", "decoder.layers.56.self_attn.out_proj.bias", "decoder.layers.56.self_attn_layer_norm.weight", "decoder.layers.56.self_attn_layer_norm.bias", "decoder.layers.56.fc1.weight", "decoder.layers.56.fc1.bias", "decoder.layers.56.fc2.weight", "decoder.layers.56.fc2.bias", "decoder.layers.56.final_layer_norm.weight", "decoder.layers.56.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.57.flat_param_0": {"names": ["decoder.layers.57.self_attn.qkv_proj.weight", "decoder.layers.57.self_attn.qkv_proj.bias", "decoder.layers.57.self_attn.out_proj.weight", "decoder.layers.57.self_attn.out_proj.bias", "decoder.layers.57.self_attn_layer_norm.weight", "decoder.layers.57.self_attn_layer_norm.bias", "decoder.layers.57.fc1.weight", "decoder.layers.57.fc1.bias", "decoder.layers.57.fc2.weight", "decoder.layers.57.fc2.bias", "decoder.layers.57.final_layer_norm.weight", "decoder.layers.57.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.58.flat_param_0": {"names": ["decoder.layers.58.self_attn.qkv_proj.weight", "decoder.layers.58.self_attn.qkv_proj.bias", "decoder.layers.58.self_attn.out_proj.weight", "decoder.layers.58.self_attn.out_proj.bias", "decoder.layers.58.self_attn_layer_norm.weight", "decoder.layers.58.self_attn_layer_norm.bias", "decoder.layers.58.fc1.weight", "decoder.layers.58.fc1.bias", "decoder.layers.58.fc2.weight", "decoder.layers.58.fc2.bias", "decoder.layers.58.final_layer_norm.weight", "decoder.layers.58.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.59.flat_param_0": {"names": ["decoder.layers.59.self_attn.qkv_proj.weight", "decoder.layers.59.self_attn.qkv_proj.bias", "decoder.layers.59.self_attn.out_proj.weight", "decoder.layers.59.self_attn.out_proj.bias", "decoder.layers.59.self_attn_layer_norm.weight", "decoder.layers.59.self_attn_layer_norm.bias", "decoder.layers.59.fc1.weight", "decoder.layers.59.fc1.bias", "decoder.layers.59.fc2.weight", "decoder.layers.59.fc2.bias", "decoder.layers.59.final_layer_norm.weight", "decoder.layers.59.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.60.flat_param_0": {"names": ["decoder.layers.60.self_attn.qkv_proj.weight", "decoder.layers.60.self_attn.qkv_proj.bias", "decoder.layers.60.self_attn.out_proj.weight", "decoder.layers.60.self_attn.out_proj.bias", "decoder.layers.60.self_attn_layer_norm.weight", "decoder.layers.60.self_attn_layer_norm.bias", "decoder.layers.60.fc1.weight", "decoder.layers.60.fc1.bias", "decoder.layers.60.fc2.weight", "decoder.layers.60.fc2.bias", "decoder.layers.60.final_layer_norm.weight", "decoder.layers.60.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.61.flat_param_0": {"names": ["decoder.layers.61.self_attn.qkv_proj.weight", "decoder.layers.61.self_attn.qkv_proj.bias", "decoder.layers.61.self_attn.out_proj.weight", "decoder.layers.61.self_attn.out_proj.bias", "decoder.layers.61.self_attn_layer_norm.weight", "decoder.layers.61.self_attn_layer_norm.bias", "decoder.layers.61.fc1.weight", "decoder.layers.61.fc1.bias", "decoder.layers.61.fc2.weight", "decoder.layers.61.fc2.bias", "decoder.layers.61.final_layer_norm.weight", "decoder.layers.61.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.62.flat_param_0": {"names": ["decoder.layers.62.self_attn.qkv_proj.weight", "decoder.layers.62.self_attn.qkv_proj.bias", "decoder.layers.62.self_attn.out_proj.weight", "decoder.layers.62.self_attn.out_proj.bias", "decoder.layers.62.self_attn_layer_norm.weight", "decoder.layers.62.self_attn_layer_norm.bias", "decoder.layers.62.fc1.weight", "decoder.layers.62.fc1.bias", "decoder.layers.62.fc2.weight", "decoder.layers.62.fc2.bias", "decoder.layers.62.final_layer_norm.weight", "decoder.layers.62.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.63.flat_param_0": {"names": ["decoder.layers.63.self_attn.qkv_proj.weight", "decoder.layers.63.self_attn.qkv_proj.bias", "decoder.layers.63.self_attn.out_proj.weight", "decoder.layers.63.self_attn.out_proj.bias", "decoder.layers.63.self_attn_layer_norm.weight", "decoder.layers.63.self_attn_layer_norm.bias", "decoder.layers.63.fc1.weight", "decoder.layers.63.fc1.bias", "decoder.layers.63.fc2.weight", "decoder.layers.63.fc2.bias", "decoder.layers.63.final_layer_norm.weight", "decoder.layers.63.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.64.flat_param_0": {"names": ["decoder.layers.64.self_attn.qkv_proj.weight", "decoder.layers.64.self_attn.qkv_proj.bias", "decoder.layers.64.self_attn.out_proj.weight", "decoder.layers.64.self_attn.out_proj.bias", "decoder.layers.64.self_attn_layer_norm.weight", "decoder.layers.64.self_attn_layer_norm.bias", "decoder.layers.64.fc1.weight", "decoder.layers.64.fc1.bias", "decoder.layers.64.fc2.weight", "decoder.layers.64.fc2.bias", "decoder.layers.64.final_layer_norm.weight", "decoder.layers.64.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.65.flat_param_0": {"names": ["decoder.layers.65.self_attn.qkv_proj.weight", "decoder.layers.65.self_attn.qkv_proj.bias", "decoder.layers.65.self_attn.out_proj.weight", "decoder.layers.65.self_attn.out_proj.bias", "decoder.layers.65.self_attn_layer_norm.weight", "decoder.layers.65.self_attn_layer_norm.bias", "decoder.layers.65.fc1.weight", "decoder.layers.65.fc1.bias", "decoder.layers.65.fc2.weight", "decoder.layers.65.fc2.bias", "decoder.layers.65.final_layer_norm.weight", "decoder.layers.65.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.66.flat_param_0": {"names": ["decoder.layers.66.self_attn.qkv_proj.weight", "decoder.layers.66.self_attn.qkv_proj.bias", "decoder.layers.66.self_attn.out_proj.weight", "decoder.layers.66.self_attn.out_proj.bias", "decoder.layers.66.self_attn_layer_norm.weight", "decoder.layers.66.self_attn_layer_norm.bias", "decoder.layers.66.fc1.weight", "decoder.layers.66.fc1.bias", "decoder.layers.66.fc2.weight", "decoder.layers.66.fc2.bias", "decoder.layers.66.final_layer_norm.weight", "decoder.layers.66.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.67.flat_param_0": {"names": ["decoder.layers.67.self_attn.qkv_proj.weight", "decoder.layers.67.self_attn.qkv_proj.bias", "decoder.layers.67.self_attn.out_proj.weight", "decoder.layers.67.self_attn.out_proj.bias", "decoder.layers.67.self_attn_layer_norm.weight", "decoder.layers.67.self_attn_layer_norm.bias", "decoder.layers.67.fc1.weight", "decoder.layers.67.fc1.bias", "decoder.layers.67.fc2.weight", "decoder.layers.67.fc2.bias", "decoder.layers.67.final_layer_norm.weight", "decoder.layers.67.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.68.flat_param_0": {"names": ["decoder.layers.68.self_attn.qkv_proj.weight", "decoder.layers.68.self_attn.qkv_proj.bias", "decoder.layers.68.self_attn.out_proj.weight", "decoder.layers.68.self_attn.out_proj.bias", "decoder.layers.68.self_attn_layer_norm.weight", "decoder.layers.68.self_attn_layer_norm.bias", "decoder.layers.68.fc1.weight", "decoder.layers.68.fc1.bias", "decoder.layers.68.fc2.weight", "decoder.layers.68.fc2.bias", "decoder.layers.68.final_layer_norm.weight", "decoder.layers.68.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.69.flat_param_0": {"names": ["decoder.layers.69.self_attn.qkv_proj.weight", "decoder.layers.69.self_attn.qkv_proj.bias", "decoder.layers.69.self_attn.out_proj.weight", "decoder.layers.69.self_attn.out_proj.bias", "decoder.layers.69.self_attn_layer_norm.weight", "decoder.layers.69.self_attn_layer_norm.bias", "decoder.layers.69.fc1.weight", "decoder.layers.69.fc1.bias", "decoder.layers.69.fc2.weight", "decoder.layers.69.fc2.bias", "decoder.layers.69.final_layer_norm.weight", "decoder.layers.69.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.70.flat_param_0": {"names": ["decoder.layers.70.self_attn.qkv_proj.weight", "decoder.layers.70.self_attn.qkv_proj.bias", "decoder.layers.70.self_attn.out_proj.weight", "decoder.layers.70.self_attn.out_proj.bias", "decoder.layers.70.self_attn_layer_norm.weight", "decoder.layers.70.self_attn_layer_norm.bias", "decoder.layers.70.fc1.weight", "decoder.layers.70.fc1.bias", "decoder.layers.70.fc2.weight", "decoder.layers.70.fc2.bias", "decoder.layers.70.final_layer_norm.weight", "decoder.layers.70.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.71.flat_param_0": {"names": ["decoder.layers.71.self_attn.qkv_proj.weight", "decoder.layers.71.self_attn.qkv_proj.bias", "decoder.layers.71.self_attn.out_proj.weight", "decoder.layers.71.self_attn.out_proj.bias", "decoder.layers.71.self_attn_layer_norm.weight", "decoder.layers.71.self_attn_layer_norm.bias", "decoder.layers.71.fc1.weight", "decoder.layers.71.fc1.bias", "decoder.layers.71.fc2.weight", "decoder.layers.71.fc2.bias", "decoder.layers.71.final_layer_norm.weight", "decoder.layers.71.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.72.flat_param_0": {"names": ["decoder.layers.72.self_attn.qkv_proj.weight", "decoder.layers.72.self_attn.qkv_proj.bias", "decoder.layers.72.self_attn.out_proj.weight", "decoder.layers.72.self_attn.out_proj.bias", "decoder.layers.72.self_attn_layer_norm.weight", "decoder.layers.72.self_attn_layer_norm.bias", "decoder.layers.72.fc1.weight", "decoder.layers.72.fc1.bias", "decoder.layers.72.fc2.weight", "decoder.layers.72.fc2.bias", "decoder.layers.72.final_layer_norm.weight", "decoder.layers.72.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.73.flat_param_0": {"names": ["decoder.layers.73.self_attn.qkv_proj.weight", "decoder.layers.73.self_attn.qkv_proj.bias", "decoder.layers.73.self_attn.out_proj.weight", "decoder.layers.73.self_attn.out_proj.bias", "decoder.layers.73.self_attn_layer_norm.weight", "decoder.layers.73.self_attn_layer_norm.bias", "decoder.layers.73.fc1.weight", "decoder.layers.73.fc1.bias", "decoder.layers.73.fc2.weight", "decoder.layers.73.fc2.bias", "decoder.layers.73.final_layer_norm.weight", "decoder.layers.73.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.74.flat_param_0": {"names": ["decoder.layers.74.self_attn.qkv_proj.weight", "decoder.layers.74.self_attn.qkv_proj.bias", "decoder.layers.74.self_attn.out_proj.weight", "decoder.layers.74.self_attn.out_proj.bias", "decoder.layers.74.self_attn_layer_norm.weight", "decoder.layers.74.self_attn_layer_norm.bias", "decoder.layers.74.fc1.weight", "decoder.layers.74.fc1.bias", "decoder.layers.74.fc2.weight", "decoder.layers.74.fc2.bias", "decoder.layers.74.final_layer_norm.weight", "decoder.layers.74.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.75.flat_param_0": {"names": ["decoder.layers.75.self_attn.qkv_proj.weight", "decoder.layers.75.self_attn.qkv_proj.bias", "decoder.layers.75.self_attn.out_proj.weight", "decoder.layers.75.self_attn.out_proj.bias", "decoder.layers.75.self_attn_layer_norm.weight", "decoder.layers.75.self_attn_layer_norm.bias", "decoder.layers.75.fc1.weight", "decoder.layers.75.fc1.bias", "decoder.layers.75.fc2.weight", "decoder.layers.75.fc2.bias", "decoder.layers.75.final_layer_norm.weight", "decoder.layers.75.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.76.flat_param_0": {"names": ["decoder.layers.76.self_attn.qkv_proj.weight", "decoder.layers.76.self_attn.qkv_proj.bias", "decoder.layers.76.self_attn.out_proj.weight", "decoder.layers.76.self_attn.out_proj.bias", "decoder.layers.76.self_attn_layer_norm.weight", "decoder.layers.76.self_attn_layer_norm.bias", "decoder.layers.76.fc1.weight", "decoder.layers.76.fc1.bias", "decoder.layers.76.fc2.weight", "decoder.layers.76.fc2.bias", "decoder.layers.76.final_layer_norm.weight", "decoder.layers.76.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.77.flat_param_0": {"names": ["decoder.layers.77.self_attn.qkv_proj.weight", "decoder.layers.77.self_attn.qkv_proj.bias", "decoder.layers.77.self_attn.out_proj.weight", "decoder.layers.77.self_attn.out_proj.bias", "decoder.layers.77.self_attn_layer_norm.weight", "decoder.layers.77.self_attn_layer_norm.bias", "decoder.layers.77.fc1.weight", "decoder.layers.77.fc1.bias", "decoder.layers.77.fc2.weight", "decoder.layers.77.fc2.bias", "decoder.layers.77.final_layer_norm.weight", "decoder.layers.77.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.78.flat_param_0": {"names": ["decoder.layers.78.self_attn.qkv_proj.weight", "decoder.layers.78.self_attn.qkv_proj.bias", "decoder.layers.78.self_attn.out_proj.weight", "decoder.layers.78.self_attn.out_proj.bias", "decoder.layers.78.self_attn_layer_norm.weight", "decoder.layers.78.self_attn_layer_norm.bias", "decoder.layers.78.fc1.weight", "decoder.layers.78.fc1.bias", "decoder.layers.78.fc2.weight", "decoder.layers.78.fc2.bias", "decoder.layers.78.final_layer_norm.weight", "decoder.layers.78.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.79.flat_param_0": {"names": ["decoder.layers.79.self_attn.qkv_proj.weight", "decoder.layers.79.self_attn.qkv_proj.bias", "decoder.layers.79.self_attn.out_proj.weight", "decoder.layers.79.self_attn.out_proj.bias", "decoder.layers.79.self_attn_layer_norm.weight", "decoder.layers.79.self_attn_layer_norm.bias", "decoder.layers.79.fc1.weight", "decoder.layers.79.fc1.bias", "decoder.layers.79.fc2.weight", "decoder.layers.79.fc2.bias", "decoder.layers.79.final_layer_norm.weight", "decoder.layers.79.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.80.flat_param_0": {"names": ["decoder.layers.80.self_attn.qkv_proj.weight", "decoder.layers.80.self_attn.qkv_proj.bias", "decoder.layers.80.self_attn.out_proj.weight", "decoder.layers.80.self_attn.out_proj.bias", "decoder.layers.80.self_attn_layer_norm.weight", "decoder.layers.80.self_attn_layer_norm.bias", "decoder.layers.80.fc1.weight", "decoder.layers.80.fc1.bias", "decoder.layers.80.fc2.weight", "decoder.layers.80.fc2.bias", "decoder.layers.80.final_layer_norm.weight", "decoder.layers.80.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.81.flat_param_0": {"names": ["decoder.layers.81.self_attn.qkv_proj.weight", "decoder.layers.81.self_attn.qkv_proj.bias", "decoder.layers.81.self_attn.out_proj.weight", "decoder.layers.81.self_attn.out_proj.bias", "decoder.layers.81.self_attn_layer_norm.weight", "decoder.layers.81.self_attn_layer_norm.bias", "decoder.layers.81.fc1.weight", "decoder.layers.81.fc1.bias", "decoder.layers.81.fc2.weight", "decoder.layers.81.fc2.bias", "decoder.layers.81.final_layer_norm.weight", "decoder.layers.81.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.82.flat_param_0": {"names": ["decoder.layers.82.self_attn.qkv_proj.weight", "decoder.layers.82.self_attn.qkv_proj.bias", "decoder.layers.82.self_attn.out_proj.weight", "decoder.layers.82.self_attn.out_proj.bias", "decoder.layers.82.self_attn_layer_norm.weight", "decoder.layers.82.self_attn_layer_norm.bias", "decoder.layers.82.fc1.weight", "decoder.layers.82.fc1.bias", "decoder.layers.82.fc2.weight", "decoder.layers.82.fc2.bias", "decoder.layers.82.final_layer_norm.weight", "decoder.layers.82.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.83.flat_param_0": {"names": ["decoder.layers.83.self_attn.qkv_proj.weight", "decoder.layers.83.self_attn.qkv_proj.bias", "decoder.layers.83.self_attn.out_proj.weight", "decoder.layers.83.self_attn.out_proj.bias", "decoder.layers.83.self_attn_layer_norm.weight", "decoder.layers.83.self_attn_layer_norm.bias", "decoder.layers.83.fc1.weight", "decoder.layers.83.fc1.bias", "decoder.layers.83.fc2.weight", "decoder.layers.83.fc2.bias", "decoder.layers.83.final_layer_norm.weight", "decoder.layers.83.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.84.flat_param_0": {"names": ["decoder.layers.84.self_attn.qkv_proj.weight", "decoder.layers.84.self_attn.qkv_proj.bias", "decoder.layers.84.self_attn.out_proj.weight", "decoder.layers.84.self_attn.out_proj.bias", "decoder.layers.84.self_attn_layer_norm.weight", "decoder.layers.84.self_attn_layer_norm.bias", "decoder.layers.84.fc1.weight", "decoder.layers.84.fc1.bias", "decoder.layers.84.fc2.weight", "decoder.layers.84.fc2.bias", "decoder.layers.84.final_layer_norm.weight", "decoder.layers.84.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.85.flat_param_0": {"names": ["decoder.layers.85.self_attn.qkv_proj.weight", "decoder.layers.85.self_attn.qkv_proj.bias", "decoder.layers.85.self_attn.out_proj.weight", "decoder.layers.85.self_attn.out_proj.bias", "decoder.layers.85.self_attn_layer_norm.weight", "decoder.layers.85.self_attn_layer_norm.bias", "decoder.layers.85.fc1.weight", "decoder.layers.85.fc1.bias", "decoder.layers.85.fc2.weight", "decoder.layers.85.fc2.bias", "decoder.layers.85.final_layer_norm.weight", "decoder.layers.85.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.86.flat_param_0": {"names": ["decoder.layers.86.self_attn.qkv_proj.weight", "decoder.layers.86.self_attn.qkv_proj.bias", "decoder.layers.86.self_attn.out_proj.weight", "decoder.layers.86.self_attn.out_proj.bias", "decoder.layers.86.self_attn_layer_norm.weight", "decoder.layers.86.self_attn_layer_norm.bias", "decoder.layers.86.fc1.weight", "decoder.layers.86.fc1.bias", "decoder.layers.86.fc2.weight", "decoder.layers.86.fc2.bias", "decoder.layers.86.final_layer_norm.weight", "decoder.layers.86.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.87.flat_param_0": {"names": ["decoder.layers.87.self_attn.qkv_proj.weight", "decoder.layers.87.self_attn.qkv_proj.bias", "decoder.layers.87.self_attn.out_proj.weight", "decoder.layers.87.self_attn.out_proj.bias", "decoder.layers.87.self_attn_layer_norm.weight", "decoder.layers.87.self_attn_layer_norm.bias", "decoder.layers.87.fc1.weight", "decoder.layers.87.fc1.bias", "decoder.layers.87.fc2.weight", "decoder.layers.87.fc2.bias", "decoder.layers.87.final_layer_norm.weight", "decoder.layers.87.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.88.flat_param_0": {"names": ["decoder.layers.88.self_attn.qkv_proj.weight", "decoder.layers.88.self_attn.qkv_proj.bias", "decoder.layers.88.self_attn.out_proj.weight", "decoder.layers.88.self_attn.out_proj.bias", "decoder.layers.88.self_attn_layer_norm.weight", "decoder.layers.88.self_attn_layer_norm.bias", "decoder.layers.88.fc1.weight", "decoder.layers.88.fc1.bias", "decoder.layers.88.fc2.weight", "decoder.layers.88.fc2.bias", "decoder.layers.88.final_layer_norm.weight", "decoder.layers.88.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.89.flat_param_0": {"names": ["decoder.layers.89.self_attn.qkv_proj.weight", "decoder.layers.89.self_attn.qkv_proj.bias", "decoder.layers.89.self_attn.out_proj.weight", "decoder.layers.89.self_attn.out_proj.bias", "decoder.layers.89.self_attn_layer_norm.weight", "decoder.layers.89.self_attn_layer_norm.bias", "decoder.layers.89.fc1.weight", "decoder.layers.89.fc1.bias", "decoder.layers.89.fc2.weight", "decoder.layers.89.fc2.bias", "decoder.layers.89.final_layer_norm.weight", "decoder.layers.89.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.90.flat_param_0": {"names": ["decoder.layers.90.self_attn.qkv_proj.weight", "decoder.layers.90.self_attn.qkv_proj.bias", "decoder.layers.90.self_attn.out_proj.weight", "decoder.layers.90.self_attn.out_proj.bias", "decoder.layers.90.self_attn_layer_norm.weight", "decoder.layers.90.self_attn_layer_norm.bias", "decoder.layers.90.fc1.weight", "decoder.layers.90.fc1.bias", "decoder.layers.90.fc2.weight", "decoder.layers.90.fc2.bias", "decoder.layers.90.final_layer_norm.weight", "decoder.layers.90.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.91.flat_param_0": {"names": ["decoder.layers.91.self_attn.qkv_proj.weight", "decoder.layers.91.self_attn.qkv_proj.bias", "decoder.layers.91.self_attn.out_proj.weight", "decoder.layers.91.self_attn.out_proj.bias", "decoder.layers.91.self_attn_layer_norm.weight", "decoder.layers.91.self_attn_layer_norm.bias", "decoder.layers.91.fc1.weight", "decoder.layers.91.fc1.bias", "decoder.layers.91.fc2.weight", "decoder.layers.91.fc2.bias", "decoder.layers.91.final_layer_norm.weight", "decoder.layers.91.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.92.flat_param_0": {"names": ["decoder.layers.92.self_attn.qkv_proj.weight", "decoder.layers.92.self_attn.qkv_proj.bias", "decoder.layers.92.self_attn.out_proj.weight", "decoder.layers.92.self_attn.out_proj.bias", "decoder.layers.92.self_attn_layer_norm.weight", "decoder.layers.92.self_attn_layer_norm.bias", "decoder.layers.92.fc1.weight", "decoder.layers.92.fc1.bias", "decoder.layers.92.fc2.weight", "decoder.layers.92.fc2.bias", "decoder.layers.92.final_layer_norm.weight", "decoder.layers.92.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.93.flat_param_0": {"names": ["decoder.layers.93.self_attn.qkv_proj.weight", "decoder.layers.93.self_attn.qkv_proj.bias", "decoder.layers.93.self_attn.out_proj.weight", "decoder.layers.93.self_attn.out_proj.bias", "decoder.layers.93.self_attn_layer_norm.weight", "decoder.layers.93.self_attn_layer_norm.bias", "decoder.layers.93.fc1.weight", "decoder.layers.93.fc1.bias", "decoder.layers.93.fc2.weight", "decoder.layers.93.fc2.bias", "decoder.layers.93.final_layer_norm.weight", "decoder.layers.93.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.94.flat_param_0": {"names": ["decoder.layers.94.self_attn.qkv_proj.weight", "decoder.layers.94.self_attn.qkv_proj.bias", "decoder.layers.94.self_attn.out_proj.weight", "decoder.layers.94.self_attn.out_proj.bias", "decoder.layers.94.self_attn_layer_norm.weight", "decoder.layers.94.self_attn_layer_norm.bias", "decoder.layers.94.fc1.weight", "decoder.layers.94.fc1.bias", "decoder.layers.94.fc2.weight", "decoder.layers.94.fc2.bias", "decoder.layers.94.final_layer_norm.weight", "decoder.layers.94.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}, "decoder.layers.95.flat_param_0": {"names": ["decoder.layers.95.self_attn.qkv_proj.weight", "decoder.layers.95.self_attn.qkv_proj.bias", "decoder.layers.95.self_attn.out_proj.weight", "decoder.layers.95.self_attn.out_proj.bias", "decoder.layers.95.self_attn_layer_norm.weight", "decoder.layers.95.self_attn_layer_norm.bias", "decoder.layers.95.fc1.weight", "decoder.layers.95.fc1.bias", "decoder.layers.95.fc2.weight", "decoder.layers.95.fc2.bias", "decoder.layers.95.final_layer_norm.weight", "decoder.layers.95.final_layer_norm.bias"], "shapes": [[4608, 12288], [4608], [12288, 1536], [12288], [12288], [12288], [6144, 12288], [6144], [12288, 6144], [12288], [12288], [12288]], "numels": [56623104, 4608, 18874368, 12288, 12288, 12288, 75497472, 6144, 75497472, 12288, 12288, 12288]}}
\ No newline at end of file

examples/tutorial/opt/inference/script/process-opt-175b/unflat.sh

+#!/usr/bin/env sh
+
+for i in $(seq 0 7); do
+    python convert_ckpt.py $1 $2 ${i} &
+done
+
+wait $(jobs -p)

examples/tutorial/opt/inference/script/processing_ckpt_66b.py

+import os
+import torch
+from multiprocessing import Pool
+
+# download pytorch model ckpt in https://huggingface.co/facebook/opt-66b/tree/main
+# you can use whether wget or git lfs
+
+path = "/path/to/your/ckpt"
+new_path = "/path/to/the/processed/ckpt/"
+
+assert os.path.isdir(path)
+files = []
+for filename in os.listdir(path):
+    filepath = os.path.join(path, filename)
+    if os.path.isfile(filepath):
+        files.append(filepath)
+
+with Pool(14) as pool:
+    ckpts = pool.map(torch.load, files)
+
+restored = {}
+for ckpt in ckpts:
+    for k,v in ckpt.items():
+        if(k[0] == 'm'):
+            k = k[6:]      
+        if(k == "lm_head.weight"):
+            k = "head.dense.weight"
+        if(k == "decoder.final_layer_norm.weight"):
+            k = "decoder.layer_norm.weight"
+        if(k == "decoder.final_layer_norm.bias"):
+            k = "decoder.layer_norm.bias"
+        restored[k] = v
+restored["decoder.version"] = "0.0"
+
+
+split_num = len(restored.keys()) // 60
+count = 0
+file_count = 1
+tmp = {}
+for k,v in restored.items():
+    print(k)
+    tmp[k] = v
+    count = count + 1    
+    if(count == split_num):
+        filename = str(file_count) + "-restored.pt"
+        torch.save(tmp, os.path.join(new_path, filename))
+        file_count = file_count + 1
+        count = 0
+        tmp = {}
+
+filename = str(file_count) + "-restored.pt"
+torch.save(tmp, os.path.join(new_path, filename))
+
+
+