Spaces:
Running on Zero
Running on Zero
Upload folder using huggingface_hub
Browse files- app.py +38 -3
- requirements.txt +1 -1
app.py
CHANGED
|
@@ -5,6 +5,7 @@ import spaces
|
|
| 5 |
import torch
|
| 6 |
import torch.nn.functional as F
|
| 7 |
import gradio as gr
|
|
|
|
| 8 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 9 |
|
| 10 |
MODEL_ID = "WhirlwindAI/Arithmetic-SLM"
|
|
@@ -16,12 +17,46 @@ CTX_LEN = 2048
|
|
| 16 |
|
| 17 |
STOP_STRINGS = [IM_END, IM_START]
|
| 18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
# ---------------------------------------------------------------------------
|
| 20 |
# Load model + tokenizer once, at module scope, moved eagerly to CUDA so
|
| 21 |
# ZeroGPU can pack the weights and stream them into VRAM on the first call.
|
| 22 |
-
# The model uses
|
| 23 |
-
#
|
| 24 |
-
#
|
| 25 |
# ---------------------------------------------------------------------------
|
| 26 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 27 |
model = AutoModelForCausalLM.from_pretrained(
|
|
|
|
| 5 |
import torch
|
| 6 |
import torch.nn.functional as F
|
| 7 |
import gradio as gr
|
| 8 |
+
import transformers.modeling_utils as _mu
|
| 9 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 10 |
|
| 11 |
MODEL_ID = "WhirlwindAI/Arithmetic-SLM"
|
|
|
|
| 17 |
|
| 18 |
STOP_STRINGS = [IM_END, IM_START]
|
| 19 |
|
| 20 |
+
# ---------------------------------------------------------------------------
|
| 21 |
+
# Compatibility shim: the model ships custom code (TinyGPTForCausalLM) written
|
| 22 |
+
# for transformers 4.57.x, whose custom `tie_weights(self)` takes no kwargs.
|
| 23 |
+
# transformers 5.x calls `self.tie_weights(recompute_mapping=False)` from
|
| 24 |
+
# `init_weights`, which raises TypeError on the custom override. We can't pin
|
| 25 |
+
# transformers 4.x (its huggingface-hub<1.0 conflicts with the Gradio 6 SDK's
|
| 26 |
+
# hub>=1.2), so instead we make the base `tie_weights` swallow the new kwarg
|
| 27 |
+
# for any subclass override that doesn't accept it. This does not change the
|
| 28 |
+
# author's inference logic — the model has tied embeddings either way.
|
| 29 |
+
# ---------------------------------------------------------------------------
|
| 30 |
+
_orig_init_weights = _mu.PreTrainedModel.init_weights
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def _safe_init_weights(self, *args, **kwargs):
|
| 34 |
+
_tie = self.tie_weights
|
| 35 |
+
|
| 36 |
+
def _tolerant_tie(*a, **k):
|
| 37 |
+
try:
|
| 38 |
+
return _tie(*a, **k)
|
| 39 |
+
except TypeError:
|
| 40 |
+
return _tie()
|
| 41 |
+
|
| 42 |
+
self.tie_weights = _tolerant_tie
|
| 43 |
+
try:
|
| 44 |
+
return _orig_init_weights(self, *args, **kwargs)
|
| 45 |
+
finally:
|
| 46 |
+
try:
|
| 47 |
+
del self.tie_weights
|
| 48 |
+
except AttributeError:
|
| 49 |
+
pass
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
_mu.PreTrainedModel.init_weights = _safe_init_weights
|
| 53 |
+
|
| 54 |
# ---------------------------------------------------------------------------
|
| 55 |
# Load model + tokenizer once, at module scope, moved eagerly to CUDA so
|
| 56 |
# ZeroGPU can pack the weights and stream them into VRAM on the first call.
|
| 57 |
+
# The model uses the pure-torch attention backend (config:
|
| 58 |
+
# attention_backend="torch", torch_fallback=True) so no flash kernels are
|
| 59 |
+
# needed at runtime.
|
| 60 |
# ---------------------------------------------------------------------------
|
| 61 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 62 |
model = AutoModelForCausalLM.from_pretrained(
|
requirements.txt
CHANGED
|
@@ -1,2 +1,2 @@
|
|
| 1 |
-
transformers
|
| 2 |
safetensors
|
|
|
|
| 1 |
+
transformers
|
| 2 |
safetensors
|