Skip to content

Fix up the mammoth max length issue. #335

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

Merged
sam-scale merged 4 commits into from
Oct 20, 2023
Merged

Fix up the mammoth max length issue. #335

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6b06342
Fix up the mammoth max length issue.
sam-scale Oct 19, 2023
289e0ef
Merge branch 'main' of ssh://github.com/scaleapi/llm-engine into ss/m…
sam-scale Oct 19, 2023
c90ba47
clean up typing
sam-scale Oct 20, 2023
cc7d2cd
Merge branch 'main' of ssh://github.com/scaleapi/llm-engine into ss/m…
sam-scale Oct 20, 2023
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
35 changes: 28 additions & 7 deletions model-engine/model_engine_server/domain/use_cases/llm_model_endpoint_use_cases.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,6 +124,12 @@
"mistral-7b-instruct": "mistralai/Mistral-7B-Instruct-v0.1",
"falcon-180b": "tiiuae/falcon-180B",
"falcon-180b-chat": "tiiuae/falcon-180B-chat",
"code-llama-7b": "codellama/CodeLlama-7b-hf",
"code-llama-13b": "codellama/CodeLlama-13b-hf",
"code-llama-34b": "codellama/CodeLlama-34b-hf",
"mammoth-coder-llama-2-7b": "TIGER-Lab/MAmmoTH-Coder-7B",
"mammoth-coder-llama-2-13b": "TIGER-Lab/MAmmoTH-Coder-13B",
"mammoth-coder-llama-2-34b": "TIGER-Lab/MAmmoTH-Coder-34B",
},
LLMInferenceFramework.LIGHTLLM: {
"llama-7b": "decapoda-research/llama-7b-hf",
Expand All @@ -143,6 +149,20 @@
LLMInferenceFramework.LIGHTLLM: [],
}

# We need a dict where if we need to override we can
# NOTE: These are in *descending* order of priority. e.g. if you see 'mammoth-coder'
# you'll use that override and not listen to the 'llama-2' override
_VLLM_MODEL_LENGTH_OVERRIDES: Dict[str, Dict[str, int]] = {
"mammoth-coder": {"max_model_len": 16384, "max_num_batched_tokens": 16384},
# Based on config here: https://huggingface.co/TIGER-Lab/MAmmoTH-Coder-7B/blob/main/config.json#L12
# Can also see 13B, 34B there too
"code-llama": {"max_model_len": 16384, "max_num_batched_tokens": 16384},
# Based on config here: https://huggingface.co/codellama/CodeLlama-7b-hf/blob/main/config.json#L12
# Can also see 13B, 34B there too
"llama-2": {"max_model_len": 4096, "max_num_batched_tokens": 4096},
"mistral": {"max_model_len": 8000, "max_num_batched_tokens": 8000},
}


NUM_DOWNSTREAM_REQUEST_RETRIES = 80 # has to be high enough so that the retries take the 5 minutes
DOWNSTREAM_REQUEST_TIMEOUT_SECONDS = 5 * 60 # 5 minutes
Expand Down Expand Up @@ -514,13 +534,14 @@ async def create_vllm_bundle(
):
command = []

max_num_batched_tokens = 2560 # vLLM's default
max_model_len = None
if "llama-2" in model_name:
max_num_batched_tokens = 4096 # Need to be bigger than model's context window
if "mistral" in model_name:
max_num_batched_tokens = 8000
max_model_len = 8000
max_num_batched_tokens: int = 2560 # vLLM's default
max_model_len: Optional[int] = None

for key, value in _VLLM_MODEL_LENGTH_OVERRIDES.items():
if key in model_name:
max_model_len = value["max_model_len"]
max_num_batched_tokens = value["max_num_batched_tokens"]
break

subcommands = []
if checkpoint_path is not None:
Expand Down