SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Support dequantizing GGUF FP16 format (#31783)

Browse Source 
* support gguf fp16

* support gguf bf16 with pytorch

* add gguf f16 test

* remove bf16
This commit is contained in:
Penut Chen
2024-07-24 23:59:59 +08:00
committed by GitHub
parent af0e4b7b37
commit 1c122a46dc
2 changed files with 17 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
tests/quantization/ggml/test_ggml.py
View File

@@ -33,6 +33,7 @@ class GgufIntegrationTests(unittest.TestCase):
mistral_model_id = "TheBloke/Mistral-7B-Instruct-v0.2-GGUF"
qwen2_model_id = "Qwen/Qwen1.5-0.5B-Chat-GGUF"
llama3_model_id = "NousResearch/Meta-Llama-3-8B-GGUF"
tinyllama_model_id = "PenutChen/TinyLlama-1.1B-Chat-v1.0-GGUF"
q4_0_gguf_model_id = "tinyllama-1.1b-chat-v1.0.Q4_0.gguf"
q4_k_gguf_model_id = "tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf"
@@ -45,6 +46,7 @@ class GgufIntegrationTests(unittest.TestCase):
q4_0_mistral_model_id = "mistral-7b-instruct-v0.2.Q4_0.gguf"
q4_0_qwen2_model_id = "qwen1_5-0_5b-chat-q4_0.gguf"
q4_llama3_model_id = "Meta-Llama-3-8B-Q4_K_M.gguf"
f16_tinyllama_model_id = "TinyLlama-1.1B-Chat-v1.0.FP16.gguf"
example_text = "Hello"
@@ -149,6 +151,18 @@ class GgufIntegrationTests(unittest.TestCase):
EXPECTED_TEXT = "Hello, World!\n\n5. Use a library"
self.assertEqual(tokenizer.decode(out[0], skip_special_tokens=True), EXPECTED_TEXT)
def test_f16(self):
tokenizer = AutoTokenizer.from_pretrained(self.tinyllama_model_id, gguf_file=self.f16_tinyllama_model_id)
model = AutoModelForCausalLM.from_pretrained(
self.tinyllama_model_id, gguf_file=self.f16_tinyllama_model_id
).to(torch_device)
text = tokenizer(self.example_text, return_tensors="pt").to(torch_device)
out = model.generate(**text, max_new_tokens=10)
EXPECTED_TEXT = "Hello, World!\n\n5. Node.js"
self.assertEqual(tokenizer.decode(out[0], skip_special_tokens=True), EXPECTED_TEXT)
def test_mistral_q4_0(self):
tokenizer = AutoTokenizer.from_pretrained(self.mistral_model_id, gguf_file=self.q4_0_mistral_model_id)
model = AutoModelForCausalLM.from_pretrained(