EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Improve gemma and documentations (#278)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-03-11 04:43:39 -07:00
committed by GitHub
parent 89885b31ef
commit faba293a0d
9 changed files with 56 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
python/sglang/backend/runtime_endpoint.py
View File

@@ -21,7 +21,9 @@ class RuntimeEndpoint(BaseBackend):
self.verify = verify
res = http_request(
self.base_url + "/get_model_info", auth_token=self.auth_token, verify=self.verify
self.base_url + "/get_model_info",
auth_token=self.auth_token,
verify=self.verify,
)
assert res.status_code == 200
self.model_info = res.json()
@@ -41,7 +43,7 @@ class RuntimeEndpoint(BaseBackend):
self.base_url + "/generate",
json={"text": prefix_str, "sampling_params": {"max_new_tokens": 0}},
auth_token=self.auth_token,
verify=self.verify
verify=self.verify,
)
assert res.status_code == 200
@@ -50,7 +52,7 @@ class RuntimeEndpoint(BaseBackend):
self.base_url + "/generate",
json={"text": s.text_, "sampling_params": {"max_new_tokens": 0}},
auth_token=self.auth_token,
verify=self.verify
verify=self.verify,
)
assert res.status_code == 200
@@ -58,7 +60,10 @@ class RuntimeEndpoint(BaseBackend):
data = {"text": s.text_, "sampling_params": {"max_new_tokens": 0}}
self._add_images(s, data)
res = http_request(
self.base_url + "/generate", json=data, auth_token=self.auth_token, verify=self.verify
self.base_url + "/generate",
json=data,
auth_token=self.auth_token,
verify=self.verify,
)
assert res.status_code == 200
@@ -90,7 +95,10 @@ class RuntimeEndpoint(BaseBackend):
self._add_images(s, data)
res = http_request(
self.base_url + "/generate", json=data, auth_token=self.auth_token, verify=self.verify
self.base_url + "/generate",
json=data,
auth_token=self.auth_token,
verify=self.verify,
)
obj = res.json()
comp = obj["text"]
@@ -129,7 +137,7 @@ class RuntimeEndpoint(BaseBackend):
json=data,
stream=True,
auth_token=self.auth_token,
verify=self.verify
verify=self.verify,
)
pos = 0
@@ -161,7 +169,10 @@ class RuntimeEndpoint(BaseBackend):
data = {"text": s.text_, "sampling_params": {"max_new_tokens": 0}}
self._add_images(s, data)
res = http_request(
self.base_url + "/generate", json=data, auth_token=self.auth_token, verify=self.verify
self.base_url + "/generate",
json=data,
auth_token=self.auth_token,
verify=self.verify,
)
assert res.status_code == 200
prompt_len = res.json()["meta_info"]["prompt_tokens"]
@@ -175,7 +186,10 @@ class RuntimeEndpoint(BaseBackend):
}
self._add_images(s, data)
res = http_request(
self.base_url + "/generate", json=data, auth_token=self.auth_token, verify=self.verify
self.base_url + "/generate",
json=data,
auth_token=self.auth_token,
verify=self.verify,
)
assert res.status_code == 200
obj = res.json()
@@ -192,7 +206,7 @@ class RuntimeEndpoint(BaseBackend):
self.base_url + "/concate_and_append_request",
json={"src_rids": src_rids, "dst_rid": dst_rid},
auth_token=self.auth_token,
verify=self.verify
verify=self.verify,
)
assert res.status_code == 200

2
python/sglang/srt/layers/token_attention.py
View File

@@ -4,8 +4,8 @@
import torch
import triton
import triton.language as tl
from sglang.srt.utils import wrap_kernel_launcher
from sglang.srt.managers.router.model_runner import global_server_args
from sglang.srt.utils import wrap_kernel_launcher
if global_server_args.attention_reduce_in_fp32:
REDUCE_TRITON_TYPE = tl.float32

18
python/sglang/srt/models/gemma.py
View File

@@ -7,7 +7,7 @@ import torch
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.radix_attention import RadixAttention
from torch import nn
from transformers import GemmaConfig
from transformers import PretrainedConfig
from vllm.config import LoRAConfig
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import GeluAndMul
@@ -136,7 +136,7 @@ class GemmaAttention(nn.Module):
class GemmaDecoderLayer(nn.Module):
def __init__(
self,
config: GemmaConfig,
config: PretrainedConfig,
layer_id: int = 0,
linear_method: Optional[LinearMethodBase] = None,
) -> None:
@@ -190,7 +190,7 @@ class GemmaDecoderLayer(nn.Module):
class GemmaModel(nn.Module):
def __init__(
self,
config: GemmaConfig,
config: PretrainedConfig,
linear_method: Optional[LinearMethodBase] = None,
) -> None:
super().__init__()
@@ -213,12 +213,12 @@ class GemmaModel(nn.Module):
input_ids: torch.Tensor,
positions: torch.Tensor,
input_metadata: InputMetadata,
skip_embed: bool = False,
input_embeds: torch.Tensor = None,
) -> torch.Tensor:
if not skip_embed:
if input_embeds is None:
hidden_states = self.embed_tokens(input_ids)
else:
hidden_states = input_ids
hidden_states = input_embeds
# Normalize the embedding by sqrt(hidden_size)
hidden_states *= self.config.hidden_size**0.5
@@ -262,7 +262,7 @@ class GemmaForCausalLM(nn.Module):
def __init__(
self,
config: GemmaConfig,
config: PretrainedConfig,
linear_method: Optional[LinearMethodBase] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
@@ -279,9 +279,9 @@ class GemmaForCausalLM(nn.Module):
input_ids: torch.Tensor,
positions: torch.Tensor,
input_metadata: InputMetadata,
skip_embed: bool = False,
input_embeds: torch.Tensor = None,
) -> torch.Tensor:
hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
return self.logits_processor(
input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
)

4
python/sglang/srt/models/llava.py
View File

@@ -233,9 +233,7 @@ class LlavaLlamaForCausalLM(nn.Module):
input_ids, positions, input_metadata, input_embeds=input_embeds
)
elif input_metadata.forward_mode == ForwardMode.DECODE:
return self.language_model(
input_ids, positions, input_metadata
)
return self.language_model(input_ids, positions, input_metadata)
def load_weights(
self,

2
python/sglang/srt/server.py
View File

@@ -550,6 +550,7 @@ class Runtime:
tp_size: int = 1,
model_mode: List[str] = (),
schedule_heuristic: str = "lpm",
attention_reduce_in_fp32: bool = False,
random_seed: int = 42,
log_level: str = "error",
port: Optional[int] = None,
@@ -572,6 +573,7 @@ class Runtime:
tp_size=tp_size,
model_mode=model_mode,
schedule_heuristic=schedule_heuristic,
attention_reduce_in_fp32=attention_reduce_in_fp32,
random_seed=random_seed,
log_level=log_level,
)

12
python/sglang/srt/server_args.py
View File

@@ -21,6 +21,7 @@ class ServerArgs:
model_mode: List[str] = ()
schedule_heuristic: str = "lpm"
schedule_conservativeness: float = 1.0
attention_reduce_in_fp32: bool = False
random_seed: int = 42
stream_interval: int = 8
disable_log_stats: bool = False
@@ -28,7 +29,6 @@ class ServerArgs:
log_level: str = "info"
disable_regex_jump_forward: bool = False
disable_disk_cache: bool = False
attention_reduce_in_fp32: bool = False
def __post_init__(self):
if self.tokenizer_path is None:
@@ -157,6 +157,11 @@ class ServerArgs:
default=ServerArgs.random_seed,
help="Random seed.",
)
parser.add_argument(
"--attention-reduce-in-fp32",
action="store_true",
help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16.",
)
parser.add_argument(
"--stream-interval",
type=int,
@@ -190,11 +195,6 @@ class ServerArgs:
action="store_true",
help="Disable disk cache to avoid possible crashes related to file system or high concurrency.",
)
parser.add_argument(
"--attention-reduce-in-fp32",
action="store_true",
help="Cast the intermidiate attention results to fp32 to avoid possible crashes related to fp16.",
)
@classmethod
def from_cli_args(cls, args: argparse.Namespace):

4
python/sglang/utils.py
View File

@@ -97,7 +97,9 @@ def http_request(url, json=None, stream=False, auth_token=None, verify=None):
"Content-Type": "application/json",
"Authentication": f"Bearer {auth_token}",
}
return requests.post(url, json=json, stream=True, headers=headers, verify=verify)
return requests.post(
url, json=json, stream=True, headers=headers, verify=verify
)
else:
req = urllib.request.Request(url)
req.add_header("Content-Type", "application/json; charset=utf-8")