cristiano.hoshikawa/nemo_guardrails_oci_generative_ai
1
0
Fork 0
mirror of https://github.com/hoshikawa2/nemo_guardrails_oci_generative_ai.git synced 2026-07-09 16:04:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Modify OCI configuration and disable Hugging Face model

Browse Source 
Updated OCI profile and provider settings, disabled OpenClaw tools, and commented out Hugging Face model code.
This commit is contained in:
Cristiano Hoshikawa
2026-04-28 07:48:31 -03:00
committed by GitHub
parent 5e36118a15
commit 5906c7865f
1 changed files with 46 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
files/oci_openai_proxy.py
View File

@@ -20,7 +20,7 @@ import torch
# ============================================================ # ============================================================
OCI_CONFIG_FILE = os.getenv("OCI_CONFIG_FILE", os.path.expanduser("~/.oci/config")) OCI_CONFIG_FILE = os.getenv("OCI_CONFIG_FILE", os.path.expanduser("~/.oci/config"))
OCI_PROFILE = os.getenv("OCI_PROFILE", "DEFAULT") OCI_PROFILE = os.getenv("OCI_PROFILE", "LATINOAMERICA-Chicago")
OCI_COMPARTMENT_ID = os.getenv("OCI_COMPARTMENT_ID", "<YOUR_COMPARTMENT_ID>") OCI_COMPARTMENT_ID = os.getenv("OCI_COMPARTMENT_ID", "<YOUR_COMPARTMENT_ID>")
OCI_GENAI_ENDPOINT = os.getenv( OCI_GENAI_ENDPOINT = os.getenv(
"OCI_GENAI_ENDPOINT", "OCI_GENAI_ENDPOINT",
@@ -29,19 +29,19 @@ OCI_GENAI_ENDPOINT = os.getenv(
if not OCI_COMPARTMENT_ID: if not OCI_COMPARTMENT_ID:
raise RuntimeError("OCI_COMPARTMENT_ID not defined") raise RuntimeError("OCI_COMPARTMENT_ID not defined")
OPENCLAW_TOOLS_ACTIVE = True OPENCLAW_TOOLS_ACTIVE = False
HF_MODEL_NAME = os.getenv("HF_MODEL_NAME", "meta-llama/Llama-4-Maverick-17B-128E-Instruct") HF_MODEL_NAME = os.getenv("HF_MODEL_NAME", "meta-llama/Llama-4-Maverick-17B-128E-Instruct")
PROVIDER = "HUGGINGFACE" PROVIDER = "OCI"
# ============================================================ # ============================================================
# HUGGINGFACE # HUGGINGFACE
# ============================================================ # ============================================================
tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_NAME) #tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained( #model = AutoModelForCausalLM.from_pretrained(
HF_MODEL_NAME, # HF_MODEL_NAME,
torch_dtype=torch.float16, # torch_dtype=torch.float16,
device_map="auto" # device_map="auto"
) #)
def normalize_messages(messages): def normalize_messages(messages):
out = [] out = []
@@ -116,49 +116,49 @@ def call_chat(body: dict, system_prompt: str):
return call_huggingface_chat(body=body, system_prompt=system_prompt) return call_huggingface_chat(body=body, system_prompt=system_prompt)
def call_huggingface_chat(body: dict, system_prompt: str): def call_huggingface_chat(body: dict, system_prompt: str):
return None
# messages = body.get("messages", [])
# messages = normalize_messages(messages)
# prompt = build_prompt(messages, system_prompt)
messages = body.get("messages", []) # inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
messages = normalize_messages(messages)
prompt = build_prompt(messages, system_prompt)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device) # with torch.no_grad():
# temperature = float(body.get("temperature", 0.0))
# top_p = float(body.get("top_p", 1.0))
with torch.no_grad(): # gen_kwargs = {
temperature = float(body.get("temperature", 0.0)) # "max_new_tokens": int(body.get("max_tokens", 512)),
top_p = float(body.get("top_p", 1.0)) # "eos_token_id": tokenizer.eos_token_id,
# }
gen_kwargs = { # if temperature > 0:
"max_new_tokens": int(body.get("max_tokens", 512)), # gen_kwargs.update({
"eos_token_id": tokenizer.eos_token_id, # "do_sample": True,
} # "temperature": temperature,
# "top_p": top_p
# })
# else:
# gen_kwargs.update({
# "do_sample": False
# })
if temperature > 0: # outputs = model.generate(**inputs, **gen_kwargs)
gen_kwargs.update({
"do_sample": True,
"temperature": temperature,
"top_p": top_p
})
else:
gen_kwargs.update({
"do_sample": False
})
outputs = model.generate(**inputs, **gen_kwargs) # generated = tokenizer.decode(outputs[0], skip_special_tokens=True)
generated = tokenizer.decode(outputs[0], skip_special_tokens=True) # # 🔥 extrai só a resposta final
# response_text = generated[len(prompt):].strip()
# 🔥 extrai só a resposta final # return {
response_text = generated[len(prompt):].strip() # "choices": [{
# "message": {
return { # "role": "assistant",
"choices": [{ # "content": response_text
"message": { # },
"role": "assistant", # "finishReason": "stop"
"content": response_text # }]
}, # }
"finishReason": "stop"
}]
}
# ============================================================ # ============================================================
# PROMPTS to adapt for OCI # PROMPTS to adapt for OCI
# ============================================================ # ============================================================
@@ -725,7 +725,8 @@ async def chat_completions(request: Request):
chat_response = run_exec_loop(body, max_steps=10000) chat_response = run_exec_loop(body, max_steps=10000)
else: else:
# 🔥 Modo enterprise → seu agent_loop controla tools # 🔥 Modo enterprise → seu agent_loop controla tools
chat_response = agent_loop(body) # chat_response = agent_loop(body)
chat_response = call_chat(body, "")
# print("FINAL RESPONSE:", json.dumps(chat_response, indent=2)) # print("FINAL RESPONSE:", json.dumps(chat_response, indent=2))