Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

remove old openai models, group models by provider #472

Closed
wants to merge 5 commits into from
Closed

remove old openai models, group models by provider #472

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
27b0c5b
remove old openai models, group models by provider
mitya52 Dec 20, 2024
47dd9fa
do not double models
mitya52 Dec 20, 2024
a675fef
local models first
mitya52 Dec 20, 2024
dbe93a8
caps version logic for 3rdparty
mitya52 Dec 23, 2024
e720dfa
litellm non streamer patch with caps version
mitya52 Dec 23, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
73 changes: 29 additions & 44 deletions refact_known_models/passthrough.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
# refer to https://docs.litellm.ai/docs/providers/

passthrough_mini_db = {
# OpenAI models
"gpt-4o": {
"backend": "litellm",
"provider": "openai",
Expand All @@ -12,28 +13,41 @@
"pp1000t_generated": 15_000, # $15.00 / 1M tokens (2024 may)
"filter_caps": ["chat", "tools", "completion"],
},
"gpt-4-turbo": {
"gpt-4o-2024-05-13": {
"backend": "litellm",
"provider": "openai",
"tokenizer_path": "Xenova/gpt-4",
"resolve_as": "gpt-4-turbo",
"tokenizer_path": "Xenova/gpt-4o",
"resolve_as": "gpt-4o-2024-05-13",
"T": 128_000,
"T_out": 4096,
"pp1000t_prompt": 10_000,
"pp1000t_generated": 30_000, # $30.00 / 1M tokens (2024 may)
"pp1000t_prompt": 5_000,
"pp1000t_generated": 15_000, # $15.00 / 1M tokens
"filter_caps": ["chat", "tools", "completion"],
},
"gpt-3.5-turbo": {
"gpt-4o-2024-08-06": {
"backend": "litellm",
"provider": "openai",
"tokenizer_path": "Xenova/gpt-3.5-turbo-16k",
"resolve_as": "gpt-3.5-turbo-1106",
"T": 16_000,
"tokenizer_path": "Xenova/gpt-4o",
"resolve_as": "gpt-4o-2024-08-06",
"T": 128_000,
"T_out": 4096,
"pp1000t_prompt": 1000,
"pp1000t_generated": 2000,
"pp1000t_prompt": 2_500,
"pp1000t_generated": 10_000, # $15.00 / 1M tokens
"filter_caps": ["chat", "tools", "completion"]
},
"gpt-4o-mini": {
"backend": "litellm",
"provider": "openai",
"tokenizer_path": "Xenova/gpt-4o",
"resolve_as": "gpt-4o-mini-2024-07-18",
"T": 128_000,
"T_out": 4096,
"pp1000t_prompt": 150,
"pp1000t_generated": 600, # $0.60 / 1M tokens
"filter_caps": ["chat", "tools", "completion"],
},

# Anthropic models
"claude-3-5-sonnet": {
"backend": "litellm",
"provider": "anthropic",
Expand Down Expand Up @@ -78,39 +92,6 @@
"pp1000t_generated": 15_000,
"filter_caps": ["chat", "tools", "completion"],
},
"gpt-4o-2024-05-13": {
"backend": "litellm",
"provider": "openai",
"tokenizer_path": "Xenova/gpt-4o",
"resolve_as": "gpt-4o-2024-05-13",
"T": 128_000,
"T_out": 4096,
"pp1000t_prompt": 5_000,
"pp1000t_generated": 15_000, # $15.00 / 1M tokens
"filter_caps": ["chat", "tools", "completion"],
},
"gpt-4o-2024-08-06": {
"backend": "litellm",
"provider": "openai",
"tokenizer_path": "Xenova/gpt-4o",
"resolve_as": "gpt-4o-2024-08-06",
"T": 128_000,
"T_out": 4096,
"pp1000t_prompt": 2_500,
"pp1000t_generated": 10_000, # $15.00 / 1M tokens
"filter_caps": ["chat", "tools", "completion"]
},
"gpt-4o-mini": {
"backend": "litellm",
"provider": "openai",
"tokenizer_path": "Xenova/gpt-4o",
"resolve_as": "gpt-4o-mini-2024-07-18",
"T": 128_000,
"T_out": 4096,
"pp1000t_prompt": 150,
"pp1000t_generated": 600, # $0.60 / 1M tokens
"filter_caps": ["chat", "tools", "completion"],
},
"claude-3-5-sonnet-20241022": {
"backend": "litellm",
"provider": "anthropic",
Expand All @@ -122,6 +103,8 @@
"pp1000t_generated": 15_000, # $15.00 / 1M tokens (2024 oct)
"filter_caps": ["chat", "tools", "completion"],
},

# Groq models
"groq-llama-3.1-8b": {
"backend": "litellm",
"provider": "groq",
Expand Down Expand Up @@ -188,6 +171,8 @@
"pp1000t_generated": 600, # TODO: don't know the price
"filter_caps": ["chat", "completion"],
},

# Cerebras models
"cerebras-llama3.1-8b": {
"backend": "litellm",
"provider": "cerebras",
Expand Down
28 changes: 14 additions & 14 deletions refact_utils/finetune/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,20 @@ def _add_results_for_passthrough_provider(provider: str) -> None:
if 'completion' in v.get('filter_caps', []):
result['completion'].append(k)

for k, v in data.get("model_assign", {}).items():
if model_dict := [d for d in data['models'] if d['name'] == k]:
model_dict = model_dict[0]

add_result(k, model_dict)

if not model_dict.get('has_finetune'):
continue

finetune_info = model_dict.get('finetune_info', []) or []
for run in finetune_info:
val = f"{k}:{run['run_id']}:{run['checkpoint']}"
add_result(val, model_dict)

if data.get("openai_api_enable"):
_add_results_for_passthrough_provider('openai')

Expand All @@ -121,20 +135,6 @@ def _add_results_for_passthrough_provider(provider: str) -> None:
if data.get('xai_api_enable'):
_add_results_for_passthrough_provider('xai')

for k, v in data.get("model_assign", {}).items():
if model_dict := [d for d in data['models'] if d['name'] == k]:
model_dict = model_dict[0]

add_result(k, model_dict)

if not model_dict.get('has_finetune'):
continue

finetune_info = model_dict.get('finetune_info', []) or []
for run in finetune_info:
val = f"{k}:{run['run_id']}:{run['checkpoint']}"
add_result(val, model_dict)

return result


Expand Down
33 changes: 24 additions & 9 deletions refact_webgui/webgui/selfhost_fastapi_completions.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -270,7 +270,7 @@ def _caps_base_data(self) -> Dict[str, Any]:
"telemetry_basic_dest": "/stats/telemetry-basic",
"telemetry_corrected_snippets_dest": "/stats/telemetry-snippets",
"telemetry_basic_retrieve_my_own": "/stats/rh-stats",
"running_models": [r for r in [*running['completion'], *running['chat']]],
"running_models": list(set(r for r in [*running['completion'], *running['chat']])),
"code_completion_default_model": code_completion_default_model,
"multiline_code_completion_default_model": multiline_code_completion_default_model,
"code_chat_default_model": code_chat_default_model,
Expand Down Expand Up @@ -455,6 +455,8 @@ async def _generate_embeddings(self, account: str, inputs: Union[str, List[str]]

async def _embeddings_style_openai(self, post: EmbeddingsStyleOpenAI, authorization: str = Header(None)):
account = await self._account_from_bearer(authorization)
# TODO: we'll implement caps_version logic later

data = [
{
"embedding": res["embedding"],
Expand Down Expand Up @@ -511,6 +513,8 @@ def compose_usage_dict(model_dict, prompt_tokens_n, generated_tokens_n) -> Dict[
return usage_dict

_account = await self._account_from_bearer(authorization)
caps_version = self._caps_version

messages = []
for m in (i.dict() for i in post.messages):
# drop tool_calls if empty, otherwise litellm tokenizing won't work
Expand All @@ -519,6 +523,16 @@ def compose_usage_dict(model_dict, prompt_tokens_n, generated_tokens_n) -> Dict[
messages.append(m)

prefix, postfix = "data: ", "\n\n"

def _patch_caps_version(data: Dict) -> Dict:
return {
**data,
"caps_version": caps_version,
}

def _wrap_output(output: str) -> str:
return prefix + output + postfix

model_dict = self._model_assigner.models_db_with_passthrough.get(post.model, {})

async def litellm_streamer():
Expand Down Expand Up @@ -546,19 +560,19 @@ async def litellm_streamer():

except json.JSONDecodeError:
data = {"choices": [{"finish_reason": finish_reason}]}
yield prefix + json.dumps(data) + postfix
yield _wrap_output(json.dumps(_patch_caps_version(data)))

final_msg = {"choices": []}
usage_dict = compose_usage_dict(model_dict, prompt_tokens_n, generated_tokens_n)
final_msg.update(usage_dict)
yield prefix + json.dumps(final_msg) + postfix
yield _wrap_output(json.dumps(_patch_caps_version(final_msg)))

# NOTE: DONE needed by refact-lsp server
yield prefix + "[DONE]" + postfix
yield _wrap_output("[DONE]")
except BaseException as e:
err_msg = f"litellm error (1): {e}"
log(err_msg)
yield prefix + json.dumps({"error": err_msg}) + postfix
yield _wrap_output(json.dumps(_patch_caps_version({"error": err_msg})))

async def litellm_non_streamer():
generated_tokens_n = 0
Expand All @@ -584,11 +598,11 @@ async def litellm_non_streamer():
data.update(usage_dict)
except json.JSONDecodeError:
data = {"choices": [{"finish_reason": finish_reason}]}
yield json.dumps(data)
yield json.dumps(_patch_caps_version(data))
except BaseException as e:
err_msg = f"litellm error (2): {e}"
log(err_msg)
yield json.dumps({"error": err_msg})
yield json.dumps(_patch_caps_version({"error": err_msg}))

async def chat_completion_streamer():
post_url = "http://127.0.0.1:8001/v1/chat"
Expand All @@ -613,11 +627,11 @@ async def chat_completion_streamer():
data["choices"][0]["finish_reason"] = None
except json.JSONDecodeError:
data = {"choices": [{"finish_reason": finish_reason}]}
yield prefix + json.dumps(data) + postfix
yield _wrap_output(json.dumps(_patch_caps_version(data)))
except aiohttp.ClientConnectorError as e:
err_msg = f"LSP server is not ready yet: {e}"
log(err_msg)
yield prefix + json.dumps({"error": err_msg}) + postfix
yield _wrap_output(json.dumps(_patch_caps_version({"error": err_msg})))

if model_dict.get('backend') == 'litellm':
model_name = model_dict.get('resolve_as', post.model)
Expand All @@ -629,6 +643,7 @@ async def chat_completion_streamer():
prompt_tokens_n += litellm.token_counter(model_name, text=json.dumps(post.tools))
response_streamer = litellm_streamer() if post.stream else litellm_non_streamer()
else:
# TODO: unused refact-lsp logic, remove ASAP
response_streamer = chat_completion_streamer()

return StreamingResponse(response_streamer, media_type="text/event-stream")
Expand Down
Loading