set buffer size to 8192 as default, decode precision as a string, lint (pytorch#476)

* set buffer size to 8192 as default, decode precision as a string, lint

* typo

* typo

* typo

Author: mikekgfb

build/builder.py

@@ -41,7 +41,7 @@ class BuilderArgs:
     def __post_init__(self):
         if self.device is None:
             self.device = "cuda" if torch.cuda.is_available() else "cpu"
-            
+
         if not (
             (self.checkpoint_path and self.checkpoint_path.is_file())
             or (self.checkpoint_dir and self.checkpoint_dir.is_dir())
@@ -408,10 +408,10 @@ def _initialize_model(
             print(f"Time to quantize model: {time.time() - t0q:.02f} seconds")
 
         if builder_args.setup_caches:
-            # TODO: get this from args?
-            max_seq_length = 2048
             with torch.device(builder_args.device):
-                model.setup_caches(max_batch_size=1, max_seq_length=max_seq_length)
+                model.setup_caches(
+                    max_batch_size=1, max_seq_length=model.config.max_seq_length
+                )
 
         model.to(dtype=builder_args.precision)
 

build/model.py

@@ -37,6 +37,7 @@ class ModelArgs:
     multiple_of: int = 256
     ffn_dim_multiplier: Optional[int] = None
     use_tiktoken: bool = False
+    max_seq_length: int = 8192
 
     def __post_init__(self):
         if self.n_local_heads == -1:

chat_in_browser.py

@@ -20,7 +20,6 @@ def create_app(*args):
         ["python3", "generate.py", *args], stdin=subprocess.PIPE, stdout=subprocess.PIPE
     )
 
-
     @app.route("/")
     def main():
         print("Starting chat session.")
@@ -93,7 +92,7 @@ def chat():
         # Strip "Model: " from output
         model_prefix = "Model: "
         if output.startswith(model_prefix):
-            output = output[len(model_prefix):]
+            output = output[len(model_prefix) :]
 
         global convo
 

download.py

@@ -36,19 +36,16 @@ def _download_hf_snapshot(
             ignore_patterns="*safetensors*",
         )
     except HTTPError as e:
-        if e.response.status_code == 401: # Missing HuggingFace CLI login.
+        if e.response.status_code == 401:  # Missing HuggingFace CLI login.
             print(
                 "Access denied. Create a HuggingFace account and run 'pip3 install huggingface_hub' and 'huggingface-cli login' to authenticate.",
-                file=sys.stderr
+                file=sys.stderr,
             )
             exit(1)
-        elif e.response.status_code == 403: # No access to the specific model.
+        elif e.response.status_code == 403:  # No access to the specific model.
             # The error message includes a link to request access to the given model. This prints nicely and does not include
             # a traceback.
-            print(
-                str(e),
-                file=sys.stderr
-            )
+            print(str(e), file=sys.stderr)
             exit(1)
         else:
             raise e

generate.py

@@ -11,7 +11,7 @@
 import time
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Optional, Tuple, List
+from typing import List, Optional, Tuple
 
 import torch
 import torch._dynamo.config
@@ -32,6 +32,7 @@
 B_INST, E_INST = "[INST]", "[/INST]"
 B_SYS, E_SYS = "<<SYS>>", "<</SYS>>"
 
+
 class ChatFormat:
     def __init__(self, tokenizer):
         self.tokenizer = tokenizer
@@ -62,7 +63,6 @@ def encode_dialog_prompt(self, dialog) -> List[int]:
         return tokens
 
 
-
 @dataclass
 class GeneratorArgs:
     prompt: str = "torchchat is pronounced torch-chat and is so cool because"
@@ -210,11 +210,17 @@ def decode_n_tokens(
 ):
     new_tokens, new_probs = [], []
     encountered_eos = False
-    for i in range(num_new_tokens - 1): # -1 to save space to run an EoS if dont generate it naturally
+    for i in range(
+        num_new_tokens - 1
+    ):  # -1 to save space to run an EoS if dont generate it naturally
         # Actually better for Inductor to codegen attention here
         with torch.nn.attention.sdpa_kernel([torch.nn.attention.SDPBackend.MATH]):
             next_token, next_prob = decode_one_token(
-                model, cur_token.clone(), input_pos, need_probs=need_probs, **sampling_kwargs
+                model,
+                cur_token.clone(),
+                input_pos,
+                need_probs=need_probs,
+                **sampling_kwargs,
             )
             input_pos += 1
             new_tokens.append(next_token.clone())
@@ -223,15 +229,25 @@ def decode_n_tokens(
                 new_probs.append(next_prob.clone())
             cur_token = next_token.view(1, -1)
             # encountered eos
-            if (next_token.item() == eos_token_id or (eot_id is not None and next_token.item() == eot_id)):
+            if next_token.item() == eos_token_id or (
+                eot_id is not None and next_token.item() == eot_id
+            ):
                 encountered_eos = True
-                _, _ = decode_one_token(model, cur_token, input_pos, need_probs, **sampling_kwargs)
+                _, _ = decode_one_token(
+                    model, cur_token, input_pos, need_probs, **sampling_kwargs
+                )
                 input_pos += 1
                 break
     if not encountered_eos:
-        eos_token = torch.tensor([eos_token_id if eot_id is None else eot_id], dtype=cur_token.dtype, device=cur_token.device)
+        eos_token = torch.tensor(
+            [eos_token_id if eot_id is None else eot_id],
+            dtype=cur_token.dtype,
+            device=cur_token.device,
+        )
         new_tokens.append(eos_token.clone())
-        _, _ = decode_one_token(model, eos_token.view(1, -1), input_pos, need_probs, **sampling_kwargs)
+        _, _ = decode_one_token(
+            model, eos_token.view(1, -1), input_pos, need_probs, **sampling_kwargs
+        )
         input_pos += 1
 
     return new_tokens, new_probs
@@ -337,7 +353,9 @@ def generate(
         with torch.device(device):
             model.setup_caches(max_batch_size=1, max_seq_length=max_seq_length)
             if is_speculative and draft_model is not model:
-                draft_model.setup_caches(max_batch_size=1, max_seq_length=max_seq_length)
+                draft_model.setup_caches(
+                    max_batch_size=1, max_seq_length=max_seq_length
+                )
 
     # create an empty tensor of the expected final shape and
     # fill in the current tokens
@@ -366,7 +384,9 @@ def generate(
 
     num_tokens_generated = 0
     input_pos = torch.tensor([start_pos + T], device=device, dtype=torch.int)
-    accept_counts = [0] * (speculate_k + 1) # creates array of [0, 0, 0, ...] that is speculate_k + 1 long
+    accept_counts = [0] * (
+        speculate_k + 1
+    )  # creates array of [0, 0, 0, ...] that is speculate_k + 1 long
 
     if is_speculative:
         input_pos = input_pos.item()  # for speculative decoding easier to keep on host
@@ -392,12 +412,14 @@ def generate(
             max_new_tokens - 1,
             callback=callback,
             need_probs=False,
-            eos_token_id = tokenizer.eos_id() if tokenizer else 2,
-            eot_id = tokenizer.special_tokens["<|eot_id|>"] if is_llama3_model else None,
+            eos_token_id=tokenizer.eos_id() if tokenizer else 2,
+            eot_id=tokenizer.special_tokens["<|eot_id|>"] if is_llama3_model else None,
             **sampling_kwargs,
         )
         seq[T + 1 : T + 1 + len(generated_tokens)] = torch.cat(generated_tokens)
-        seq = seq[:T + 1 + len(generated_tokens)] # If we dont generate all the way to max_new_tokens slice off the extra space we allocated.
+        seq = seq[
+            : T + 1 + len(generated_tokens)
+        ]  # If we dont generate all the way to max_new_tokens slice off the extra space we allocated.
 
     generate_stats = {"accept_counts": accept_counts}
     return seq, generate_stats
@@ -410,7 +432,6 @@ def encode_tokens(tokenizer, string, bos=True, device="cpu"):
     return torch.tensor(tokens, dtype=torch.int, device=device)
 
 
-
 def get_device_info(name: str) -> str:
     import platform
     from subprocess import check_output
@@ -481,7 +502,9 @@ def _main(
     # Piggy backing off of this flag then for now to identify llama3 without prompting user.
     is_llama3_model = tokenizer_args.is_tiktoken
     if generator_args.chat_mode and is_llama3_model:
-        logging.debug("Llama3 model detected in chat mode. Using updated sentence schemas")
+        logging.debug(
+            "Llama3 model detected in chat mode. Using updated sentence schemas"
+        )
 
     builder_args.setup_caches = False
     model = _initialize_model(builder_args, quantize, tokenizer)
@@ -534,20 +557,29 @@ def _main(
         if generator_args.compile_prefill:
             prefill = torch.compile(prefill, fullgraph=True, dynamic=True)
 
-    system_prompt=None
+    system_prompt = None
     # Set up our max_seq_length
     if generator_args.chat_mode:
-        max_seq_length = 2048
-        print(f"Entering Chat Mode. Will continue chatting back and forth with the language model until the models max context length of {max_seq_length} tokens is hit or until the user says /bye")
-        system_prompt = input("System Prompt [Optional]: ")
+        max_seq_length = model.config.max_seq_length
+        print(
+            f"Entering Chat Mode. Will continue chatting back and forth with the language model until the models max context length of {max_seq_length} tokens is hit or until the user says /bye"
+        )
+        get_system_prompt = input(
+            "Do you want to enter a system prompt? Enter y for yes and anything else for no. \n"
+        )
+        if get_system_prompt == "y" or get_system_prompt == "Y":
+            system_prompt = input("What is your system prompt? \n")
         if is_llama3_model:
             chat_formatter = ChatFormat(tokenizer)
     else:
-        max_seq_length = min(encoded.size(0) + generator_args.max_new_tokens, model.config.block_size)
-
+        max_seq_length = min(
+            encoded.size(0) + generator_args.max_new_tokens, model.config.block_size
+        )
 
     max_seq_length = (
-        max_seq_length + speculate_k + 1 if draft_model is not None else max_seq_length
+        max_seq_length + speculative_builder_args.speculate_k + 1
+        if draft_model is not None
+        else max_seq_length
     )
 
     aggregate_metrics = {
@@ -557,39 +589,59 @@ def _main(
     start = -1 if generator_args.compile else 0
     start_pos = 0
 
-
     # arbitrarily large number as chat mode goes until max_seq length or user exits
     num_samples = generator_args.num_samples if not generator_args.chat_mode else 100000
-    i = -1  # long loop and Im scared someone will add a continue in it, so start at -1 and increment at the start
-    while (i < num_samples):
+    i = (
+        -1
+    )  # long loop and Im scared someone will add a continue in it, so start at -1 and increment at the start
+    while i < num_samples:
         i += 1
         device_sync(device=builder_args.device)
         if i >= 0 and generator_args.chat_mode:
             prompt = input("User: ")
-            if (prompt == "/bye"):
+            if prompt == "/bye":
                 print("Exiting Chat.\n")
                 break
             if not is_llama3_model:
                 if system_prompt:
                     prompt = f"{B_INST} {B_SYS}\n{system_prompt.strip()}\n{E_SYS}\n\n{prompt.strip} {E_INST}"
-                    system_prompt = None # can only provide system prompt on first interaction
+                    system_prompt = (
+                        None  # can only provide system prompt on first interaction
+                    )
                 else:
                     prompt = f"{B_INST} {prompt.strip()} {E_INST}"
                 encoded = encode_tokens(
                     tokenizer, prompt, bos=True, device=builder_args.device
                 )
             else:
-                if system_prompt:
-                    encoded = chat_formatter.encode_dialog_prompt([{"role" : "system", "content" : system_prompt}, {"role" : "user", "content" : prompt}])
+                if system_prompt is not None:
+                    encoded = chat_formatter.encode_dialog_prompt(
+                        [
+                            {"role": "system", "content": system_prompt},
+                            {"role": "user", "content": prompt},
+                        ]
+                    )
                     system_prompt = None
-                elif(i == 0):
-                    encoded = chat_formatter.encode_dialog_prompt([{"role" : "user", "content" : prompt}])
+                elif i == 0:
+                    encoded = chat_formatter.encode_dialog_prompt(
+                        [{"role": "user", "content": prompt}]
+                    )
                 else:
-                    encoded = chat_formatter.encode_message({"role" : "user", "content" : prompt})
-                    encoded.extend(chat_formatter.encode_header({"role": "assistant", "content": ""}))
-                encoded = torch.tensor(encoded, dtype=torch.int, device=builder_args.device)
-            if (encoded.size(0) + start_pos > max_seq_length):
-                print("This prompt would take us past the max_seq_length. Ending Conversation.")
+                    encoded = chat_formatter.encode_message(
+                        {"role": "user", "content": prompt}
+                    )
+                    encoded.extend(
+                        chat_formatter.encode_header(
+                            {"role": "assistant", "content": ""}
+                        )
+                    )
+                encoded = torch.tensor(
+                    encoded, dtype=torch.int, device=builder_args.device
+                )
+            if encoded.size(0) + start_pos > max_seq_length:
+                print(
+                    "This prompt would take us past the max_seq_length. Ending Conversation."
+                )
                 break
 
         if generator_args.chat_mode and i >= 0:
@@ -604,12 +656,17 @@ def callback(
             ):
                 if done_generating:
                     return
-                buffer.append(tokenizer.decode([period_id] + x.tolist())[1:]) # I think this results in the first output token being dropped from the display which is wrong.
+                buffer.append(
+                    tokenizer.decode([period_id] + x.tolist())[1:]
+                )  # I think this results in the first output token being dropped from the display which is wrong.
                 if x.item() == tokenizer.eos_id():
                     done_generating = True
-                if (is_llama3_model and x.item() == tokenizer.special_tokens["<|eot_id|>"]):
+                if (
+                    is_llama3_model
+                    and x.item() == tokenizer.special_tokens["<|eot_id|>"]
+                ):
                     done_generating = True
-                    buffer = buffer[:-1] # drop the eot_id from the output buffer
+                    buffer = buffer[:-1]  # drop the eot_id from the output buffer
                 if len(buffer) == 4 or done_generating:
                     print("".join(buffer), end="", flush=True)
                     buffer.clear()
@@ -672,7 +729,7 @@ def callback(x):
         )
         logging.debug(f"Bandwidth achieved: {model_size * tokens_sec / 1e9:.02f} GB/s")
 
-        if (start_pos >= max_seq_length):
+        if start_pos >= max_seq_length:
             print("Max Sequence Length Reached. Ending Conversation.")
             break
 

quantize.py

@@ -15,7 +15,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from build.utils import find_multiple, get_precision, use_et_backend
+from build.utils import find_multiple, get_precision, name_to_dtype, use_et_backend
 
 
 #########################################################################
@@ -97,11 +97,14 @@ def quantized_model(self) -> nn.Module:
 
 
 class PrecisionHandler(QuantHandler):
-    def __init__(self, model: nn.Module, device="cpu", tokenizer=None, **kwargs):
+    def __init__(self, model: nn.Module, device="cpu", tokenizer=None, *, dtype):
         self.model_ = model
         self.device = device
         self.tokenizer = tokenizer
-        self.kwargs = kwargs
+
+        if isinstance(dtype, str):
+            dtype = name_to_dtype(dtype)
+        self.dtype = dtype
 
     def create_quantized_state_dict(self) -> Dict:  # "StateDict"
         pass
@@ -110,7 +113,7 @@ def convert_for_runtime(self) -> nn.Module:
         pass
 
     def quantized_model(self) -> nn.Module:
-        return self.model_.to(device=self.device, **self.kwargs)
+        return self.model_.to(device=self.device, dtype=self.dtype)
 
 
 #########################################################################