Commit
·
1175344
1
Parent(s):
b416f51
- hf_backend.py +24 -6
hf_backend.py
CHANGED
|
@@ -1,4 +1,3 @@
|
|
| 1 |
-
# hf_backend.py
|
| 2 |
import time, logging, json, asyncio
|
| 3 |
from contextlib import nullcontext
|
| 4 |
from typing import Any, Dict, AsyncIterable, Tuple
|
|
@@ -33,16 +32,38 @@ except Exception as e:
|
|
| 33 |
load_error = f"Failed to load tokenizer: {e}"
|
| 34 |
logger.exception(load_error)
|
| 35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
def _pick_cpu_dtype() -> torch.dtype:
|
| 37 |
try:
|
| 38 |
-
if
|
| 39 |
-
logger.info("[dtype]
|
| 40 |
return torch.bfloat16
|
| 41 |
except Exception as e:
|
| 42 |
logger.warning(f"[dtype] BF16 probe failed: {e}")
|
| 43 |
logger.info("[dtype] fallback -> torch.float32")
|
| 44 |
return torch.float32
|
| 45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
_MODEL_CACHE: Dict[tuple[str, torch.dtype], AutoModelForCausalLM] = {}
|
| 47 |
|
| 48 |
def _get_model(device: str, dtype: torch.dtype) -> Tuple[AutoModelForCausalLM, torch.dtype]:
|
|
@@ -153,12 +174,10 @@ class HFChatBackend(ChatBackend):
|
|
| 153 |
zero_client.HEADERS["X-IP-Token"] = x_ip_token
|
| 154 |
logger.info("[req] injected X-IP-Token into ZeroGPU headers")
|
| 155 |
|
| 156 |
-
# Build prompt (pass tools to template)
|
| 157 |
if hasattr(tokenizer, "apply_chat_template") and getattr(tokenizer, "chat_template", None):
|
| 158 |
try:
|
| 159 |
prompt = tokenizer.apply_chat_template(
|
| 160 |
messages,
|
| 161 |
-
#tools=tools,
|
| 162 |
tokenize=False,
|
| 163 |
add_generation_prompt=True,
|
| 164 |
)
|
|
@@ -212,7 +231,6 @@ class HFChatBackend(ChatBackend):
|
|
| 212 |
logger.info(f"[gen] text len={len(text)}\n{_snippet(text, 1200)}")
|
| 213 |
return text
|
| 214 |
|
| 215 |
-
# Offload heavy work to a worker thread so asyncio heartbeats continue
|
| 216 |
if spaces:
|
| 217 |
@spaces.GPU(duration=120)
|
| 218 |
def run_once_sync(prompt: str) -> str:
|
|
|
|
|
|
|
| 1 |
import time, logging, json, asyncio
|
| 2 |
from contextlib import nullcontext
|
| 3 |
from typing import Any, Dict, AsyncIterable, Tuple
|
|
|
|
| 32 |
load_error = f"Failed to load tokenizer: {e}"
|
| 33 |
logger.exception(load_error)
|
| 34 |
|
| 35 |
+
|
| 36 |
+
def probe_bf16_runtime() -> bool:
|
| 37 |
+
"""Check if BF16 is both reported and actually used in ops on CPU."""
|
| 38 |
+
if not (hasattr(torch, "cpu") and hasattr(torch.cpu, "is_bf16_supported")):
|
| 39 |
+
return False
|
| 40 |
+
if not torch.cpu.is_bf16_supported():
|
| 41 |
+
return False
|
| 42 |
+
try:
|
| 43 |
+
a = torch.randn(16, 16, dtype=torch.bfloat16)
|
| 44 |
+
b = torch.randn(16, 16, dtype=torch.bfloat16)
|
| 45 |
+
c = a @ b
|
| 46 |
+
return c.dtype == torch.bfloat16
|
| 47 |
+
except Exception:
|
| 48 |
+
return False
|
| 49 |
+
|
| 50 |
+
|
| 51 |
def _pick_cpu_dtype() -> torch.dtype:
|
| 52 |
try:
|
| 53 |
+
if probe_bf16_runtime():
|
| 54 |
+
logger.info("[dtype] Verified BF16 execution on CPU -> torch.bfloat16")
|
| 55 |
return torch.bfloat16
|
| 56 |
except Exception as e:
|
| 57 |
logger.warning(f"[dtype] BF16 probe failed: {e}")
|
| 58 |
logger.info("[dtype] fallback -> torch.float32")
|
| 59 |
return torch.float32
|
| 60 |
|
| 61 |
+
|
| 62 |
+
# Log CPU dtype capability at startup
|
| 63 |
+
CPU_DTYPE = _pick_cpu_dtype()
|
| 64 |
+
logger.info(f"[init] Default CPU dtype = {CPU_DTYPE}")
|
| 65 |
+
|
| 66 |
+
|
| 67 |
_MODEL_CACHE: Dict[tuple[str, torch.dtype], AutoModelForCausalLM] = {}
|
| 68 |
|
| 69 |
def _get_model(device: str, dtype: torch.dtype) -> Tuple[AutoModelForCausalLM, torch.dtype]:
|
|
|
|
| 174 |
zero_client.HEADERS["X-IP-Token"] = x_ip_token
|
| 175 |
logger.info("[req] injected X-IP-Token into ZeroGPU headers")
|
| 176 |
|
|
|
|
| 177 |
if hasattr(tokenizer, "apply_chat_template") and getattr(tokenizer, "chat_template", None):
|
| 178 |
try:
|
| 179 |
prompt = tokenizer.apply_chat_template(
|
| 180 |
messages,
|
|
|
|
| 181 |
tokenize=False,
|
| 182 |
add_generation_prompt=True,
|
| 183 |
)
|
|
|
|
| 231 |
logger.info(f"[gen] text len={len(text)}\n{_snippet(text, 1200)}")
|
| 232 |
return text
|
| 233 |
|
|
|
|
| 234 |
if spaces:
|
| 235 |
@spaces.GPU(duration=120)
|
| 236 |
def run_once_sync(prompt: str) -> str:
|