Skip to content

Commit

Permalink
Add StreamingLLM support to studio2 chat (#2060)
Browse files Browse the repository at this point in the history 
* Streaming LLM

* Update precision and add gpu support

* (studio2) Separate weights generation for quantization support

* Adapt prompt changes to studio flow

* Remove outdated flag from llm compile flags.

* (studio2) use turbine vmfbRunner

* tweaks to prompts

* Update CPU path and llm api test.

* Change device in test to cpu.

* Fixes to runner, device names, vmfb mgmt

* Use small test without external weights.
  • Loading branch information
@monorimet
monorimet committed Feb 12, 2024
1 parent 01575a8 commit 25312cd
Show file tree
Hide file tree
Showing 4 changed files with 218 additions and 163 deletions.
210 changes: 156 additions & 54 deletions apps/shark_studio/api/llm.py
View file
Original file line number Diff line number Diff line change
@@ -1,10 +1,9 @@
from turbine_models.custom_models import stateless_llama
from turbine_models.model_runner import vmfbRunner
from turbine_models.gen_external_params.gen_external_params import gen_external_params
import time
from shark.iree_utils.compile_utils import (
get_iree_compiled_module,
load_vmfb_using_mmap,
)
from apps.shark_studio.web.utils.file_utils import get_resource_path
from shark.iree_utils.compile_utils import compile_module_to_flatbuffer
from apps.shark_studio.web.utils import get_resource_path
import iree.runtime as ireert
from itertools import chain
import gc
Expand All @@ -16,61 +15,128 @@
"llama2_7b": {
"initializer": stateless_llama.export_transformer_model,
"hf_model_name": "meta-llama/Llama-2-7b-chat-hf",
"compile_flags": ["--iree-opt-const-expr-hoisting=False"],
"stop_token": 2,
"max_tokens": 4096,
"system_prompt": """<s>[INST] <<SYS>>Be concise. You are a helpful, respectful and honest assistant. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. <</SYS>>""",
},
"Trelis/Llama-2-7b-chat-hf-function-calling-v2": {
"initializer": stateless_llama.export_transformer_model,
"hf_model_name": "Trelis/Llama-2-7b-chat-hf-function-calling-v2",
"compile_flags": ["--iree-opt-const-expr-hoisting=False"],
"stop_token": 2,
"max_tokens": 4096,
"system_prompt": """<s>[INST] <<SYS>>Be concise. You are a helpful, respectful and honest assistant. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. <</SYS>>""",
},
"TinyPixel/small-llama2": {
"initializer": stateless_llama.export_transformer_model,
"hf_model_name": "TinyPixel/small-llama2",
"compile_flags": ["--iree-opt-const-expr-hoisting=True"],
"stop_token": 2,
"max_tokens": 1024,
"system_prompt": """<s>[INST] <<SYS>>Be concise. You are a helpful, respectful and honest assistant. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. <</SYS>>""",
},
}

B_INST, E_INST = "[INST]", "[/INST]"
B_SYS, E_SYS = "<s>", "</s>"

DEFAULT_CHAT_SYS_PROMPT = """<s>[INST] <<SYS>>
Be concise. You are a helpful, respectful and honest assistant. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.\n <</SYS>>\n\n
"""


def append_user_prompt(history, input_prompt):
user_prompt = f"{B_INST} {input_prompt} {E_INST}"
history += user_prompt
return history


class LanguageModel:
def __init__(
self,
model_name,
hf_auth_token=None,
device=None,
precision="fp32",
quantization="int4",
precision="",
external_weights=None,
use_system_prompt=True,
streaming_llm=False,
):
print(llm_model_map[model_name])
self.hf_model_name = llm_model_map[model_name]["hf_model_name"]
self.tempfile_name = get_resource_path("llm.torch.tempfile")
self.vmfb_name = get_resource_path("llm.vmfb.tempfile")
self.device = device
self.precision = precision
self.safe_name = self.hf_model_name.strip("/").replace("/", "_")
self.max_tokens = llm_model_map[model_name]["max_tokens"]
self.iree_module_dict = None
self.device = device.split("=>")[-1].strip()
self.backend = self.device.split("://")[0]
self.driver = self.backend
if "cpu" in device:
self.device = "cpu"
self.backend = "llvm-cpu"
self.driver = "local-task"

print(f"Selected {self.backend} as IREE target backend.")
self.precision = "f32" if "cpu" in device else "f16"
self.quantization = quantization
self.safe_name = self.hf_model_name.replace("/", "_").replace("-", "_")
self.external_weight_file = None
# TODO: find a programmatic solution for model arch spec instead of hardcoding llama2
self.file_spec = "_".join(
[
self.safe_name,
self.precision,
]
)
if self.quantization != "None":
self.file_spec += "_" + self.quantization

if external_weights is not None:
self.external_weight_file = get_resource_path(
self.safe_name + "." + external_weights
self.file_spec + "." + external_weights
)

if streaming_llm:
# Add streaming suffix to file spec after setting external weights filename.
self.file_spec += "_streaming"
self.streaming_llm = streaming_llm

self.tempfile_name = get_resource_path(f"{self.file_spec}.tempfile")
# TODO: Tag vmfb with target triple of device instead of HAL backend
self.vmfb_name = get_resource_path(
f"{self.file_spec}_{self.backend}.vmfb.tempfile"
)
self.max_tokens = llm_model_map[model_name]["max_tokens"]
self.iree_module_dict = None
self.use_system_prompt = use_system_prompt
self.global_iter = 0
self.prev_token_len = 0
self.first_input = True
if self.external_weight_file is not None:
if not os.path.exists(self.external_weight_file):
print(
f"External weight file {self.external_weight_file} does not exist. Generating..."
)
gen_external_params(
hf_model_name=self.hf_model_name,
quantization=self.quantization,
weight_path=self.external_weight_file,
hf_auth_token=hf_auth_token,
precision=self.precision,
)
else:
print(
f"External weight file {self.external_weight_file} found for {self.vmfb_name}"
)
if os.path.exists(self.vmfb_name) and (
external_weights is None or os.path.exists(str(self.external_weight_file))
):
self.iree_module_dict = dict()
(
self.iree_module_dict["vmfb"],
self.iree_module_dict["config"],
self.iree_module_dict["temp_file_to_unlink"],
) = load_vmfb_using_mmap(
self.vmfb_name,
device,
device_idx=0,
rt_flags=[],
external_weight_file=self.external_weight_file,
self.runner = vmfbRunner(
device=self.driver,
vmfb_path=self.vmfb_name,
external_weight_path=self.external_weight_file,
)
if self.streaming_llm:
self.model = self.runner.ctx.modules.streaming_state_update
else:
self.model = self.runner.ctx.modules.state_update
self.tokenizer = AutoTokenizer.from_pretrained(
self.hf_model_name,
use_fast=False,
Expand All @@ -82,7 +148,9 @@ def __init__(
hf_auth_token,
compile_to="torch",
external_weights=external_weights,
external_weight_file=self.external_weight_file,
precision=self.precision,
quantization=self.quantization,
streaming_llm=self.streaming_llm,
)
with open(self.tempfile_name, "w+") as f:
f.write(self.torch_ir)
Expand All @@ -99,30 +167,50 @@ def __init__(

def compile(self) -> None:
# this comes with keys: "vmfb", "config", and "temp_file_to_unlink".
self.iree_module_dict = get_iree_compiled_module(
# ONLY architecture/api-specific compile-time flags for each backend, if needed.
# hf_model_id-specific global flags currently in model map.
flags = []
if "cpu" in self.backend:
flags.extend(
[
"--iree-global-opt-enable-quantized-matmul-reassociation",
]
)
elif self.backend == "vulkan":
flags.extend(["--iree-stream-resource-max-allocation-size=4294967296"])
flags.extend(llm_model_map[self.hf_model_name]["compile_flags"])
flatbuffer_blob = compile_module_to_flatbuffer(
self.tempfile_name,
device=self.device,
mmap=True,
frontend="torch",
external_weight_file=self.external_weight_file,
model_config_path=None,
extra_args=flags,
write_to=self.vmfb_name,
extra_args=["--iree-global-opt-enable-quantized-matmul-reassociation"],
)
# TODO: delete the temp file
self.runner = vmfbRunner(
device=self.driver,
vmfb_path=self.vmfb_name,
external_weight_path=self.external_weight_file,
)
if self.streaming_llm:
self.model = self.runner.ctx.modules.streaming_state_update
else:
self.model = self.runner.ctx.modules.state_update

def sanitize_prompt(self, prompt):
print(prompt)
if isinstance(prompt, list):
prompt = list(chain.from_iterable(prompt))
prompt = " ".join([x for x in prompt if isinstance(x, str)])
prompt = prompt.replace("\n", " ")
prompt = prompt.replace("\t", " ")
prompt = prompt.replace("\r", " ")
if self.use_system_prompt and self.global_iter == 0:
prompt = llm_model_map["llama2_7b"]["system_prompt"] + prompt
prompt += " [/INST]"
print(prompt)
return prompt
prompt = append_user_prompt(DEFAULT_CHAT_SYS_PROMPT, prompt)
print(prompt)
return prompt
else:
print(prompt)
return f"{B_INST} {prompt} {E_INST}"

def chat(self, prompt):
prompt = self.sanitize_prompt(prompt)
Expand All @@ -134,26 +222,40 @@ def format_out(results):

history = []
for iter in range(self.max_tokens):
st_time = time.time()
if iter == 0:
device_inputs = [
ireert.asdevicearray(
self.iree_module_dict["config"].device, input_tensor
)
]
token = self.iree_module_dict["vmfb"]["run_initialize"](*device_inputs)
if self.streaming_llm:
token_slice = max(self.prev_token_len - 1, 0)
input_tensor = input_tensor[:, token_slice:]
if self.streaming_llm and self.model["get_seq_step"]() > 600:
print("Evicting cache space!")
self.model["evict_kvcache_space"]()
token_len = input_tensor.shape[-1]
device_inputs = [
ireert.asdevicearray(self.runner.config.device, input_tensor)
]
if self.first_input or not self.streaming_llm:
st_time = time.time()
token = self.model["run_initialize"](*device_inputs)
total_time = time.time() - st_time
token_len += 1
self.first_input = False
else:
device_inputs = [
ireert.asdevicearray(
self.iree_module_dict["config"].device,
token,
)
]
token = self.iree_module_dict["vmfb"]["run_forward"](*device_inputs)
st_time = time.time()
token = self.model["run_cached_initialize"](*device_inputs)
total_time = time.time() - st_time
token_len += 1

total_time = time.time() - st_time
history.append(format_out(token))
yield self.tokenizer.decode(history), total_time
while format_out(token) != llm_model_map["llama2_7b"]["stop_token"]:
dec_time = time.time()
if self.streaming_llm and self.model["get_seq_step"]() > 600:
print("Evicting cache space!")
self.model["evict_kvcache_space"]()
token = self.model["run_forward"](token)
history.append(format_out(token))
total_time = time.time() - dec_time
yield self.tokenizer.decode(history), total_time

self.prev_token_len = token_len + len(history)

if format_out(token) == llm_model_map["llama2_7b"]["stop_token"]:
break
Expand Down
17 changes: 11 additions & 6 deletions apps/shark_studio/tests/api_test.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,8 @@
import logging
import unittest
import json
from apps.shark_studio.api.llm import LanguageModel
import gc

from apps.shark_studio.api.llm import LanguageModel
from apps.shark_studio.api.sd import shark_sd_fn_dict_input, view_json_file
Expand All @@ -28,12 +30,13 @@ def testSDSimple(self):
print(i)

class LLMAPITest(unittest.TestCase):
def testLLMSimple(self):
def test01_LLMSmall(self):
lm = LanguageModel(
"Trelis/Llama-2-7b-chat-hf-function-calling-v2",
"TinyPixel/small-llama2",
hf_auth_token=None,
device="cpu-task",
external_weights="safetensors",
device="cpu",
precision="fp32",
quantization="None",
)
count = 0
for msg, _ in lm.chat("hi, what are you?"):
Expand All @@ -42,9 +45,11 @@ def testLLMSimple(self):
count += 1
continue
assert (
msg.strip(" ") == "Hello"
), f"LLM API failed to return correct response, expected 'Hello', received {msg}"
msg.strip(" ") == "Turkish Turkish Turkish"
), f"LLM API failed to return correct response, expected 'Turkish Turkish Turkish', received {msg}"
break
del lm
gc.collect()


if __name__ == "__main__":
Expand Down
Loading

0 comments on commit 25312cd

Please sign in to comment.