Instructions to use mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch

SGLang

How to use mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch with Docker Model Runner:
```
docker model run hf.co/mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch
```

mlnomad commited on Apr 27

Commit

53d9696

verified ·

1 Parent(s): a931476

Fix: lazy non-persistent RoPE buffers — HF from_pretrained's meta-init was leaving persistent rope_cos/rope_sin uninitialised, producing NaN logits. Buffers are now computed on first forward.

Browse files

Files changed (1) hide show

torch_gpt.py +22 -13

torch_gpt.py CHANGED Viewed

@@ -315,17 +315,27 @@ class GELU_GPT(nn.Module):
             }
         )
-        # Precompute RoPE tables (non-trainable, stored as buffers).
-        # Flax uses rotary_seq_len = sequence_len * 10; mirror that so long
-        # contexts + parity tests both work.
-        rotary_seq_len = config.sequence_len * 10
-        cos, sin = precompute_rotary_embeddings(
-            rotary_seq_len, config.head_dim, base=config.rope_base
-        )
-        # persistent=True so HF from_pretrained (which can meta-init
-        # non-persistent buffers as NaN) restores them correctly.
-        self.register_buffer("rope_cos", cos, persistent=True)
-        self.register_buffer("rope_sin", sin, persistent=True)
     # ------------------------------------------------------------------
     # Forward
@@ -334,8 +344,7 @@ class GELU_GPT(nn.Module):
         B, T = idx.shape
         config = self.config
-        cos = self.rope_cos[:, :T].to(dtype=self.wte.weight.dtype)
-        sin = self.rope_sin[:, :T].to(dtype=self.wte.weight.dtype)
         x = self.wte(idx)
         x = rms_norm(x)

             }
         )
+        # Lazy RoPE: HF from_pretrained can leave persistent buffers
+        # uninitialised (meta-init → NaN). We instead register an empty
+        # placeholder and compute the table on the first forward.
+        self._rope_max_len = config.sequence_len * 10
+        self._rope_head_dim = config.head_dim
+        self._rope_base = config.rope_base
+        self.register_buffer("rope_cos", torch.empty(0), persistent=False)
+        self.register_buffer("rope_sin", torch.empty(0), persistent=False)
+        self._rope_initialized = False
+    def _get_rope(self, T, dtype, device):
+        if (not self._rope_initialized
+                or self.rope_cos.numel() == 0
+                or self.rope_cos.shape[1] < T):
+            cos, sin = precompute_rotary_embeddings(
+                max(T, self._rope_max_len), self._rope_head_dim, base=self._rope_base
+            )
+            self.rope_cos = cos.to(device)
+            self.rope_sin = sin.to(device)
+            self._rope_initialized = True
+        return self.rope_cos[:, :T].to(dtype), self.rope_sin[:, :T].to(dtype)
     # ------------------------------------------------------------------
     # Forward
         B, T = idx.shape
         config = self.config
+        cos, sin = self._get_rope(T, self.wte.weight.dtype, self.wte.weight.device)
         x = self.wte(idx)
         x = rms_norm(x)