Instructions to use cmpatino/nanowhale-100m-base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cmpatino/nanowhale-100m-base with Transformers:

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

pipe = pipeline("text-generation", model="cmpatino/nanowhale-100m-base", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("cmpatino/nanowhale-100m-base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("cmpatino/nanowhale-100m-base", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use cmpatino/nanowhale-100m-base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cmpatino/nanowhale-100m-base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cmpatino/nanowhale-100m-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cmpatino/nanowhale-100m-base

SGLang

How to use cmpatino/nanowhale-100m-base 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 "cmpatino/nanowhale-100m-base" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cmpatino/nanowhale-100m-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "cmpatino/nanowhale-100m-base" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cmpatino/nanowhale-100m-base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cmpatino/nanowhale-100m-base with Docker Model Runner:
```
docker model run hf.co/cmpatino/nanowhale-100m-base
```

cmpatino HF Staff commited on 4 days ago

Commit

d10cc5f

verified ·

1 Parent(s): 96016bc

fix: use nn.init.* in _init_weights + recompute freqs_cis buffer on load

Browse files

Two bugs caused gibberish generation when loading via from_pretrained:

1. _init_weights used module.weight.data.*_() which bypasses the HF
init guard (_is_hf_initialized), causing loaded pretrained weights
to be overwritten with random initialization. Fixed by using
nn.init.normal_/nn.init.zeros_/nn.init.ones_ instead.

2. freqs_cis (RoPE frequencies) is a non-persistent buffer that gets
filled with uninitialized memory when from_pretrained constructs
on meta device. Added explicit recomputation in _init_weights.

Files changed (1) hide show

modeling_deepseek_v4.py +17 -8

modeling_deepseek_v4.py CHANGED Viewed

@@ -469,21 +469,22 @@ class DeepseekV4PreTrainedModel(PreTrainedModel):
     _skip_keys_device_placement = ["past_key_values"]
     def _init_weights(self, module):
         std = self.config.initializer_range
         if isinstance(module, nn.Linear):
-            module.weight.data.normal_(mean=0.0, std=std)
             if module.bias is not None:
-                module.bias.data.zero_()
         elif isinstance(module, nn.Embedding):
-            module.weight.data.normal_(mean=0.0, std=std)
         elif isinstance(module, DeepseekV4RMSNorm):
-            module.weight.data.fill_(1.0)
         elif isinstance(module, DeepseekV4Gate):
-            module.weight.data.normal_(mean=0.0, std=std)
             if module.bias is not None:
-                module.bias.data.zero_()
         elif isinstance(module, DeepseekV4Block):
-            # Initialize HC parameters
             nn.init.normal_(module.hc_attn_fn, std=0.01)
             nn.init.normal_(module.hc_ffn_fn, std=0.01)
             nn.init.zeros_(module.hc_attn_base)
@@ -523,11 +524,19 @@ class DeepseekV4Model(DeepseekV4PreTrainedModel):
     def _init_weights(self, module):
         super()._init_weights(module)
-        # HC head initialization
         if module is self:
             nn.init.normal_(self.hc_head_fn, std=0.01)
             nn.init.zeros_(self.hc_head_base)
             nn.init.ones_(self.hc_head_scale)
     def hc_head(self, x):
         """Contract hc_mult copies to 1 for final output.

     _skip_keys_device_placement = ["past_key_values"]
     def _init_weights(self, module):
+        # Use nn.init.* (not module.weight.data.*_()) so transformers'
+        # init guard skips already-loaded params via the _is_hf_initialized flag.
         std = self.config.initializer_range
         if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, mean=0.0, std=std)
             if module.bias is not None:
+                nn.init.zeros_(module.bias)
         elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=std)
         elif isinstance(module, DeepseekV4RMSNorm):
+            nn.init.ones_(module.weight)
         elif isinstance(module, DeepseekV4Gate):
+            nn.init.normal_(module.weight, mean=0.0, std=std)
             if module.bias is not None:
+                nn.init.zeros_(module.bias)
         elif isinstance(module, DeepseekV4Block):
             nn.init.normal_(module.hc_attn_fn, std=0.01)
             nn.init.normal_(module.hc_ffn_fn, std=0.01)
             nn.init.zeros_(module.hc_attn_base)
     def _init_weights(self, module):
         super()._init_weights(module)
         if module is self:
             nn.init.normal_(self.hc_head_fn, std=0.01)
             nn.init.zeros_(self.hc_head_base)
             nn.init.ones_(self.hc_head_scale)
+            # Recompute non-persistent freqs_cis buffer. from_pretrained constructs
+            # on meta device and refills non-persistent buffers with uninitialized
+            # memory; without this, rotary positional encodings would be garbage.
+            freqs = precompute_freqs_cis(
+                self.config.qk_rope_head_dim,
+                self.config.max_position_embeddings,
+                self.config.rope_theta,
+            ).to(self.freqs_cis.device)
+            self.freqs_cis.data.copy_(freqs)
     def hc_head(self, x):
         """Contract hc_mult copies to 1 for final output.