Merge 77c46698b9 into b5d872ead0

inference/convert.py

@@ -1,9 +1,9 @@
 import os
 import shutil
-from argparse import ArgumentParser
+from parser import Parser
 from glob import glob
 from tqdm import tqdm, trange
-
+import asyncio as sync 
 import torch
 from safetensors.torch import safe_open, save_file
 
@@ -29,6 +29,32 @@ mapping = {
     "scale": ("scale", None),
 }
 
+async def set_param(param, name, i, n_local_experts, mp, state_dicts, dim):
+    new_param = param
+    if "experts" in name and "shared_experts" not in name:
+        idx = int(name.split(".")[-3])
+        if idx < i * n_local_experts or idx >= (i + 1) * n_local_experts:
+            return
+    elif dim is not None:
+        assert param.size(dim) % mp == 0
+        shard_size = param.size(dim) // mp
+        new_param = param.narrow(dim, i * shard_size, shard_size).contiguous()
+    state_dicts[i][name] = new_param
+async def inner_safe_open(name, f, state_dicts, mp, n_local_experts): 
+        if "model.layers.61" not in name:
+            param: torch.Tensor = f.get_tensor(name)
+            if name.startswith("model."):
+                name = name[len("model."):]
+            name = name.replace("self_attn", "attn")
+            name = name.replace("mlp", "ffn")
+            name = name.replace("weight_scale_inv", "scale")
+            name = name.replace("e_score_correction_bias", "bias")
+            key = name.split(".")[-2]
+            assert key in mapping
+            new_key, dim = mapping[key]
+            name = name.replace(key, new_key)
+            await sync.gather(*(set_param(param, name, i, n_local_experts, mp, state_dicts, dim) for i in range(mp)))
+                
 
 def main(hf_ckpt_path, save_path, n_experts, mp):
     """
@@ -44,53 +70,30 @@ def main(hf_ckpt_path, save_path, n_experts, mp):
         None
     """
     torch.set_num_threads(8)
-    n_local_experts = n_experts // mp
+    n_local_experts,state_dicts = n_experts // mp, [{} for _ in range(mp)]
-    state_dicts = [{} for _ in range(mp)]
+    tensor_dir, token_dir = list(glob(os.path.join(hf_ckpt_path, "*.safetensors"))),list(glob(os.path.join(hf_ckpt_path, "*token*")))
-
+    for file_path in tqdm(tensor_dir):
-    for file_path in tqdm(glob(os.path.join(hf_ckpt_path, "*.safetensors"))):
+        cm = await sync.to_thread(safe_open, file_path, framework="pt", device="cpu")
-        with safe_open(file_path, framework="pt", device="cpu") as f:
+        async with cm as f:
-            for name in f.keys():
+            await sync.gather(*(inner_safe_open(name, f, state_dicts, mp, n_local_experts) for name in f.keys()))
-                if "model.layers.61" in name:
-                    continue
-                param: torch.Tensor = f.get_tensor(name)
-                if name.startswith("model."):
-                    name = name[len("model."):]
-                name = name.replace("self_attn", "attn")
-                name = name.replace("mlp", "ffn")
-                name = name.replace("weight_scale_inv", "scale")
-                name = name.replace("e_score_correction_bias", "bias")
-                key = name.split(".")[-2]
-                assert key in mapping
-                new_key, dim = mapping[key]
-                name = name.replace(key, new_key)
-                for i in range(mp):
-                    new_param = param
-                    if "experts" in name and "shared_experts" not in name:
-                        idx = int(name.split(".")[-3])
-                        if idx < i * n_local_experts or idx >= (i + 1) * n_local_experts:
-                            continue
-                    elif dim is not None:
-                        assert param.size(dim) % mp == 0
-                        shard_size = param.size(dim) // mp
-                        new_param = param.narrow(dim, i * shard_size, shard_size).contiguous()
-                    state_dicts[i][name] = new_param
 
     os.makedirs(save_path, exist_ok=True)
 
-    for i in trange(mp):
+    await sync.gather(*(sync.to_thread(save_file(state_dicts[i], os.path.join(save_path, f"model{i}-mp{mp}.safetensors"))) for i in trange(mp)))
-        save_file(state_dicts[i], os.path.join(save_path, f"model{i}-mp{mp}.safetensors"))
+
+    async def set_file_path(file_path):
+        await sync.to_thread(shutil.copyfile, file_path, os.path.join(save_path, os.path.basename(file_path)))
+    
+    await sync.gather(*(set_file_path(file_path) for file_path in token_dir))
 
-    for file_path in glob(os.path.join(hf_ckpt_path, "*token*")):
-        new_file_path = os.path.join(save_path, os.path.basename(file_path))
-        shutil.copyfile(file_path, new_file_path)
 
 
 if __name__ == "__main__":
-    parser = ArgumentParser()
+    arg_list = [
-    parser.add_argument("--hf-ckpt-path", type=str, required=True)
+        ("--hf-ckpt-path", type:=str, required:=True),
-    parser.add_argument("--save-path", type=str, required=True)
+        ("--save-path", type:=str, required:=True),
-    parser.add_argument("--n-experts", type=int, required=True)
+        ("--n-experts", type:=int, required:=True),
-    parser.add_argument("--model-parallel", type=int, required=True)
+        ("--model-parallel", type:=int, required:=True)
-    args = parser.parse_args()
+    ]
-    assert args.n_experts % args.model_parallel == 0
+    args = Parser(arg_list).apply_args().assert_model_parallel().return_args()
-    main(args.hf_ckpt_path, args.save_path, args.n_experts, args.model_parallel)
+    sync.run(main(args.hf_ckpt_path, args.save_path, args.n_experts, args.model_parallel))

inference/fp8_cast_bf16.py

@@ -3,13 +3,41 @@ import json
 from argparse import ArgumentParser
 from glob import glob
 from tqdm import tqdm
-
+from asyncio import gather, to_thread, run
 import torch
 from safetensors.torch import load_file, save_file
-
 from kernel import weight_dequant
 
-def main(fp8_path, bf16_path):
+def inner_tensor_file(safetensor_file):
+    file_name = os.path.basename(safetensor_file)
+    current_state_dict = load_file(safetensor_file, device="cuda")
+    loaded_files[file_name] = current_state_dict
+    new_state_dict = {}
+    for weight_name, weight in current_state_dict.items():
+        if weight_name.endswith("_scale_inv"):
+            continue
+        elif weight.element_size() == 1:  # FP8 weight
+            scale_inv_name = f"{weight_name}_scale_inv"
+            try:
+                # Get scale_inv from the correct file
+                scale_inv = get_tensor(scale_inv_name)
+                fp8_weight_names.append(weight_name)
+                new_state_dict[weight_name] = weight_dequant(weight, scale_inv)
+            except KeyError:
+                print(f"Warning: Missing scale_inv tensor for {weight_name}, skipping conversion")
+                new_state_dict[weight_name] = weight
+        else:
+            new_state_dict[weight_name] = weight
+            new_safetensor_file = os.path.join(bf16_path, file_name)
+    save_file(new_state_dict, new_safetensor_file)
+    
+    # Memory management: keep only the 2 most recently used files
+    if len(loaded_files) > 2:
+        oldest_file = next(iter(loaded_files))
+        del loaded_files[oldest_file]
+        torch.cuda.empty_cache()
+
+async def main(fp8_path, bf16_path):
     """
     Converts FP8 weights to BF16 and saves the converted weights.
 
@@ -32,13 +60,11 @@ def main(fp8_path, bf16_path):
     torch.set_default_dtype(torch.bfloat16)
     os.makedirs(bf16_path, exist_ok=True)
     model_index_file = os.path.join(fp8_path, "model.safetensors.index.json")
-    with open(model_index_file, "r") as f:
+    with open(model_index_file, "r") as f: model_index = json.load(f)
-        model_index = json.load(f)
     weight_map = model_index["weight_map"]
     
     # Cache for loaded safetensor files
-    loaded_files = {}
+    loaded_files, fp8_weight_names = {}, []
-    fp8_weight_names = []
 
     # Helper function to get tensor from the correct file
     def get_tensor(tensor_name):
@@ -62,45 +88,15 @@ def main(fp8_path, bf16_path):
 
     safetensor_files = list(glob(os.path.join(fp8_path, "*.safetensors")))
     safetensor_files.sort()
-    for safetensor_file in tqdm(safetensor_files):
+    gather(*(to_thread(inner_tensor_file, safetensor_file) for safetensor_file in tqdm(safetensor_files)))
-        file_name = os.path.basename(safetensor_file)
-        current_state_dict = load_file(safetensor_file, device="cuda")
-        loaded_files[file_name] = current_state_dict
 
-        new_state_dict = {}
-        for weight_name, weight in current_state_dict.items():
-            if weight_name.endswith("_scale_inv"):
-                continue
-            elif weight.element_size() == 1:  # FP8 weight
-                scale_inv_name = f"{weight_name}_scale_inv"
-                try:
-                    # Get scale_inv from the correct file
-                    scale_inv = get_tensor(scale_inv_name)
-                    fp8_weight_names.append(weight_name)
-                    new_state_dict[weight_name] = weight_dequant(weight, scale_inv)
-                except KeyError:
-                    print(f"Warning: Missing scale_inv tensor for {weight_name}, skipping conversion")
-                    new_state_dict[weight_name] = weight
-            else:
-                new_state_dict[weight_name] = weight
-                
-        new_safetensor_file = os.path.join(bf16_path, file_name)
-        save_file(new_state_dict, new_safetensor_file)
-        
-        # Memory management: keep only the 2 most recently used files
-        if len(loaded_files) > 2:
-            oldest_file = next(iter(loaded_files))
-            del loaded_files[oldest_file]
-            torch.cuda.empty_cache()
     
     # Update model index
     new_model_index_file = os.path.join(bf16_path, "model.safetensors.index.json")
     for weight_name in fp8_weight_names:
         scale_inv_name = f"{weight_name}_scale_inv"
-        if scale_inv_name in weight_map:
+        if scale_inv_name in weight_map: weight_map.pop(scale_inv_name)
-            weight_map.pop(scale_inv_name)
+    with open(new_model_index_file, "w") as f: json.dump({"metadata": {}, "weight_map": weight_map}, f, indent=2)
-    with open(new_model_index_file, "w") as f:
-        json.dump({"metadata": {}, "weight_map": weight_map}, f, indent=2)
         
 
 if __name__ == "__main__":
@@ -108,5 +104,5 @@ if __name__ == "__main__":
     parser.add_argument("--input-fp8-hf-path", type=str, required=True)
     parser.add_argument("--output-bf16-hf-path", type=str, required=True)
     args = parser.parse_args()
-    main(args.input_fp8_hf_path, args.output_bf16_hf_path)
+    run(main(args.input_fp8_hf_path, args.output_bf16_hf_path))
     

inference/generate.py

@@ -1,13 +1,13 @@
 import os
 import json
+from parser import Parser
 from argparse import ArgumentParser
 from typing import List
-
 import torch
 import torch.distributed as dist
 from transformers import AutoTokenizer
 from safetensors.torch import load_model
-
+from asyncio import gather, to_thread, run
 from model import Transformer, ModelArgs
 
 
@@ -36,6 +36,7 @@ def generate(
     temperature: float = 1.0
 ) -> List[List[int]]:
     """
+    
     Generates new tokens based on the given prompt tokens using the specified model.
 
     Args:
@@ -47,38 +48,35 @@ def generate(
 
     Returns:
         List[List[int]]: A list of lists containing the generated tokens for each sequence.
+        
     """
     prompt_lens = [len(t) for t in prompt_tokens]
     assert max(prompt_lens) <= model.max_seq_len
     total_len = min(model.max_seq_len, max_new_tokens + max(prompt_lens))
     tokens = torch.full((len(prompt_tokens), total_len), -1, dtype=torch.long, device="cuda")
-    for i, t in enumerate(prompt_tokens):
+    for i, t in enumerate(prompt_tokens): tokens[i, :len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
-        tokens[i, :len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
     prev_pos = 0
     finished = torch.tensor([False] * len(prompt_tokens), device="cuda")
     prompt_mask = tokens != -1
-    for cur_pos in range(min(prompt_lens), total_len):
+    def inner_cur_pos():
         logits = model.forward(tokens[:, prev_pos:cur_pos], prev_pos)
-        if temperature > 0:
+        if temperature > 0: next_token = sample(logits, temperature)
-            next_token = sample(logits, temperature)
+        else: next_token = logits.argmax(dim=-1)
-        else:
-            next_token = logits.argmax(dim=-1)
         next_token = torch.where(prompt_mask[:, cur_pos], tokens[:, cur_pos], next_token)
         tokens[:, cur_pos] = next_token
         finished |= torch.logical_and(~prompt_mask[:, cur_pos], next_token == eos_id)
         prev_pos = cur_pos
-        if finished.all():
+        if finished.all(): return
-            break
+    gather(*(to_thread(cur_pos) for cur_pos in range(min(prompt_lens), total_len)))
     completion_tokens = []
     for i, toks in enumerate(tokens.tolist()):
         toks = toks[prompt_lens[i]:prompt_lens[i]+max_new_tokens]
-        if eos_id in toks:
+        if eos_id in toks: toks = toks[:toks.index(eos_id)]
-            toks = toks[:toks.index(eos_id)]
         completion_tokens.append(toks)
     return completion_tokens
 
 
-def main(
+async def main(
     ckpt_path: str,
     config: str,
     input_file: str = "",
@@ -131,8 +129,7 @@ def main(
                 objects = [None]
                 dist.broadcast_object_list(objects, 0)
                 prompt = objects[0]
-            if prompt == "/exit":
+            if prompt == "/exit": break
-                break
             elif prompt == "/clear":
                 messages.clear()
                 continue
@@ -143,8 +140,7 @@ def main(
             print(completion)
             messages.append({"role": "assistant", "content": completion})
     else:
-        with open(input_file) as f:
+        with open(input_file) as f: prompts = [line.strip() for line in f.readlines()]
-            prompts = [line.strip() for line in f.readlines()]
         assert len(prompts) <= args.max_batch_size
         prompt_tokens = [tokenizer.apply_chat_template([{"role": "user", "content": prompt}], add_generation_prompt=True) for prompt in prompts]
         completion_tokens = generate(model, prompt_tokens, max_new_tokens, tokenizer.eos_token_id, temperature)
@@ -154,8 +150,7 @@ def main(
             print("Completion:", completion)
             print()
 
-    if world_size > 1:
+    if world_size > 1: dist.destroy_process_group()
-        dist.destroy_process_group()
 
 
 if __name__ == "__main__":
@@ -173,13 +168,13 @@ if __name__ == "__main__":
     Raises:
         AssertionError: If neither input-file nor interactive mode is specified.
     """
-    parser = ArgumentParser()
+    arg_list = [   
-    parser.add_argument("--ckpt-path", type=str, required=True)
+        ("--ckpt-path", type:=str, required:=True),
-    parser.add_argument("--config", type=str, required=True)
+        ("--config", type:=str, required:=True),
-    parser.add_argument("--input-file", type=str, default="")
+        ("--input-file", type:=str, default:=""),
-    parser.add_argument("--interactive", action="store_true")
+        ("--interactive", action:="store_true"),
-    parser.add_argument("--max-new-tokens", type=int, default=200)
+        ("--max-new-tokens", type:=int, default:=200),
-    parser.add_argument("--temperature", type=float, default=0.2)
+        ("--temperature", type:=float, default:=0.2)
-    args = parser.parse_args()
+    ]
-    assert args.input_file or args.interactive
+    args = Parser(arg_list).apply_args().assert_interactive().return_args()
-    main(args.ckpt_path, args.config, args.input_file, args.interactive, args.max_new_tokens, args.temperature)
+    run(main(args.ckpt_path, args.config, args.input_file, args.interactive, args.max_new_tokens, args.temperature))

inference/parser.py

@@ -0,0 +1,18 @@
+from argparse import ArgumentParser
+
+class Parser(): 
+    def __init__(self, parser = ArgumentParser(), arg_list = []):
+        self.parser = parser
+        self.arg_list = arg_list
+    def apply_args(self):
+        for arg in self.arg_list: self.parser.add_argument(*arg)
+        return self
+    def assert_model_parallel(self):
+        assert self.return_args.n_experts % self.return_args().model_parallel == 0
+        return self
+    def assert_interactive():
+        assert self.return_args().input_file or self.return_args().interactive
+        return self
+    def return_args(self): 
+        return self.parser.parse_args()
+