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

Add gguf support for StableLM (#33793)

Browse Source 
* add stablelm gguf architecture support

* add additional quantization tests

* resolve merge conflict, add weight conversion tests for fp16
This commit is contained in:
Vladislav Bronzov
2024-10-09 12:16:13 +02:00
committed by GitHub
parent e783f12f20
commit faa0f63b93
4 changed files with 107 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
tests/quantization/ggml/test_ggml.py
View File

@@ -48,6 +48,9 @@ class GgufIntegrationTests(unittest.TestCase):
falcon7b_model_id = "xaviviro/falcon-7b-quantized-gguf"
falcon40b_model_id = "maddes8cht/tiiuae-falcon-40b-gguf"
original_flacon7b_model_id = "tiiuae/falcon-7b"
stablelm_model_id = "afrideva/stablelm-3b-4e1t-GGUF"
stablelm2_model_id = "afrideva/stablelm-2-1_6b-GGUF"
original_stablelm2_model_id = "stabilityai/stablelm-2-1_6b"
# standard quants
q4_0_gguf_model_id = "tinyllama-1.1b-chat-v1.0.Q4_0.gguf"
@@ -59,6 +62,7 @@ class GgufIntegrationTests(unittest.TestCase):
q4_k_gguf_model_id = "tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf"
q5_k_gguf_model_id = "tinyllama-1.1b-chat-v1.0.Q5_K_M.gguf"
q6_k_gguf_model_id = "tinyllama-1.1b-chat-v1.0.Q6_K.gguf"
q4_k_m_stablelm_model_id = "stablelm-3b-4e1t.q4_k_m.gguf"
# imatrix
iq1_m_gguf_model_id = "TinyLlama-1.1B-Chat-v1.0-IQ1_M.gguf"
iq1_s_gguf_model_id = "TinyLlama-1.1B-Chat-v1.0-IQ1_S.gguf"
@@ -76,6 +80,7 @@ class GgufIntegrationTests(unittest.TestCase):
q8_qwen2moe_model_id = "Qwen1.5-MoE-A2.7B_Q8_0.gguf"
q4_llama3_model_id = "Meta-Llama-3-8B-Q4_K_M.gguf"
fp16_bloom_model_id = "bloom-560m.fp16.gguf"
fp16_stablelm2_model_id = "stablelm-2-1_6b.fp16.gguf"
q8_bloom_model_id = "bloom-560m.q8_0.gguf"
f16_tinyllama_model_id = "TinyLlama-1.1B-Chat-v1.0.FP16.gguf"
q2_k_falcon7b_model_id = "falcon-7b-q2_k.gguf"
@@ -523,6 +528,75 @@ class GgufIntegrationTests(unittest.TestCase):
self.assertTrue(original_params.shape == quantized_state_dict[layer_name].shape)
torch.testing.assert_close(original_params, quantized_state_dict[layer_name])
def test_stablelm_q4_k_m(self):
model = AutoModelForCausalLM.from_pretrained(
self.stablelm_model_id,
gguf_file=self.q4_k_m_stablelm_model_id,
device_map="auto",
torch_dtype=torch.float16,
)
tokenizer = AutoTokenizer.from_pretrained(self.stablelm_model_id, gguf_file=self.q4_k_m_stablelm_model_id)
text = tokenizer(self.example_text, return_tensors="pt").to(torch_device)
out = model.generate(**text, max_new_tokens=10)
EXPECTED_TEXT = "Hello-\nI am trying to create a new user"
self.assertEqual(tokenizer.decode(out[0], skip_special_tokens=True), EXPECTED_TEXT)
def test_stablelm_fp16(self):
original_model = AutoModelForCausalLM.from_pretrained(
self.original_stablelm2_model_id,
torch_dtype=torch.float16,
)
converted_model = AutoModelForCausalLM.from_pretrained(
self.stablelm2_model_id,
gguf_file=self.fp16_stablelm2_model_id,
torch_dtype=torch.float16,
# for precise comparison it is required to use the original model config
# as quantized one is different in parameters: use_parallel_residual and use_qkv_bias
# and it highly influences on the output results
config=original_model.config,
)
tokenizer = AutoTokenizer.from_pretrained(self.stablelm2_model_id, gguf_file=self.fp16_stablelm2_model_id)
text = tokenizer(self.example_text, return_tensors="pt")
original_out = original_model.generate(**text, max_new_tokens=10)
converted_out = converted_model.generate(**text, max_new_tokens=10)
EXPECTED_TEXT = "Hello, I am a 20 year old male"
self.assertEqual(tokenizer.decode(converted_out[0], skip_special_tokens=True), EXPECTED_TEXT)
self.assertEqual(
tokenizer.decode(converted_out[0], skip_special_tokens=True),
tokenizer.decode(original_out[0], skip_special_tokens=True),
)
def test_stablelm_weights_conversion_fp16(self):
original_model = AutoModelForCausalLM.from_pretrained(
self.original_stablelm2_model_id,
device_map="auto",
torch_dtype=torch.float16,
)
converted_model = AutoModelForCausalLM.from_pretrained(
self.stablelm2_model_id,
gguf_file=self.fp16_stablelm2_model_id,
device_map="auto",
torch_dtype=torch.float16,
# for precise comparison it is required to use the original model config
# as quantized one is different in parameters: use_parallel_residual and use_qkv_bias
# and it highly influences on the output results
config=original_model.config,
)
converted_state_dict = converted_model.state_dict()
original_state_dict = original_model.state_dict()
for layer_name, original_params in original_state_dict.items():
if layer_name in converted_state_dict:
self.assertTrue(original_params.shape == converted_state_dict[layer_name].shape)
torch.testing.assert_close(original_params, converted_state_dict[layer_name])
def test_tokenization_xnli(self):
import tqdm
from datasets import load_dataset