mgoin commited on
Commit
3a3f45d
·
verified ·
1 Parent(s): b32b43c

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +187 -0
README.md ADDED
@@ -0,0 +1,187 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - fp8
4
+ - vllm
5
+ license: mit
6
+ license_link: https://huggingface.co/microsoft/Phi-3.5-mini-instruct/resolve/main/LICENSE
7
+ ---
8
+
9
+ # Phi-3.5-mini-instruct-FP8-KV
10
+
11
+ ## Model Overview
12
+ - **Model Architecture:** Phi-3.5
13
+ - **Input:** Text
14
+ - **Output:** Text
15
+ - **Model Optimizations:**
16
+ - **Weight quantization:** FP8
17
+ - **Activation quantization:** FP8
18
+ - **Intended Use Cases:** Intended for commercial and research use in English. Similarly to [Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct), this models is intended for assistant-like chat.
19
+ - **Out-of-scope:** Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English.
20
+ - **Release Date:** 8/11/2024
21
+ - **Version:** 1.1
22
+ - **License(s):** [mit](https://huggingface.co/microsoft/Phi-3.5-mini-instruct/resolve/main/LICENSE)
23
+ - **Model Developers:** Neural Magic
24
+
25
+ Quantized version of [Phi-3.5-mini-instruct](https://huggingface.co/microsoft/Phi-3.5-mini-instruct), with the new configuration files.
26
+
27
+ ### Model Optimizations
28
+
29
+ This model was obtained by quantizing the weights and activations of [Phi-3.5-mini-instruct](https://huggingface.co/microsoft/Phi-3.5-mini-instruct) to FP8 data type, ready for inference with vLLM >= 0.5.1.
30
+ This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%.
31
+
32
+ Only the weights and activations of the linear operators within transformers blocks are quantized. Symmetric per-tensor quantization is applied, in which a single linear scaling maps the FP8 representations of the quantized weights and activations.
33
+ [AutoFP8](https://github.com/neuralmagic/AutoFP8) is used for quantization with 512 sequences of UltraChat.
34
+
35
+ ## Deployment
36
+
37
+ ### Use with vLLM
38
+
39
+ This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below.
40
+
41
+ ```python
42
+ from vllm import LLM, SamplingParams
43
+ from transformers import AutoTokenizer
44
+
45
+ model_id = "neuralmagic/Phi-3.5-mini-instruct-FP8-KV"
46
+
47
+ sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=256)
48
+
49
+ tokenizer = AutoTokenizer.from_pretrained(model_id)
50
+
51
+ messages = [
52
+ {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
53
+ {"role": "user", "content": "Who are you? Remember to respond in pirate speak!"},
54
+ ]
55
+
56
+ prompts = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
57
+
58
+ llm = LLM(model=model_id, kv_cache_dtype="fp8")
59
+
60
+ outputs = llm.generate(prompts, sampling_params)
61
+
62
+ generated_text = outputs[0].outputs[0].text
63
+ print(generated_text)
64
+ ```
65
+
66
+ vLLM aslo supports OpenAI-compatible serving. See the [documentation](https://docs.vllm.ai/en/latest/) for more details.
67
+
68
+ ## Creation
69
+
70
+ This model was created by applying [LLM Compressor with calibration samples from UltraChat](https://github.com/vllm-project/llm-compressor/blob/sa/big_model_support/examples/big_model_offloading/big_model_w8a8_calibrate.py), as presented in the code snipet below.
71
+
72
+ ```python
73
+ from datasets import load_dataset
74
+ from transformers import AutoTokenizer
75
+
76
+ from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
77
+
78
+ # Select model and load it.
79
+ # Phi-3.5 is a special case for KV cache quantization because it has
80
+ # fused QKV linear layers.
81
+ MODEL_ID = "microsoft/Phi-3.5-mini-instruct"
82
+ model = SparseAutoModelForCausalLM.from_pretrained(
83
+ MODEL_ID,
84
+ device_map="auto",
85
+ torch_dtype="auto",
86
+ )
87
+ tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
88
+
89
+ # Select calibration dataset.
90
+ DATASET_ID = "HuggingFaceH4/ultrachat_200k"
91
+ DATASET_SPLIT = "train_sft"
92
+
93
+ # Select number of samples. 512 samples is a good place to start.
94
+ # Increasing the number of samples can improve accuracy.
95
+ NUM_CALIBRATION_SAMPLES = 512
96
+ MAX_SEQUENCE_LENGTH = 2048
97
+
98
+ # Load dataset and preprocess.
99
+ ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
100
+ ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
101
+
102
+
103
+ def process_and_tokenize(example):
104
+ text = tokenizer.apply_chat_template(example["messages"], tokenize=False)
105
+ return tokenizer(
106
+ text,
107
+ padding=False,
108
+ max_length=MAX_SEQUENCE_LENGTH,
109
+ truncation=True,
110
+ add_special_tokens=False,
111
+ )
112
+
113
+
114
+ ds = ds.map(process_and_tokenize, remove_columns=ds.column_names)
115
+
116
+ # Configure the quantization algorithm and scheme.
117
+ # In this case, we:
118
+ # * quantize the weights to fp8 with per-tensor scales
119
+ # * quantize the activations to fp8 with per-tensor scales
120
+ # * quantize the kv cache to fp8 with per-tensor scales
121
+ recipe = """
122
+ quant_stage:
123
+ quant_modifiers:
124
+ QuantizationModifier:
125
+ ignore: ["lm_head"]
126
+ config_groups:
127
+ group_0:
128
+ weights:
129
+ num_bits: 8
130
+ type: float
131
+ strategy: tensor
132
+ dynamic: false
133
+ symmetric: true
134
+ input_activations:
135
+ num_bits: 8
136
+ type: float
137
+ strategy: tensor
138
+ dynamic: false
139
+ symmetric: true
140
+ targets: ["Linear"]
141
+ kv_cache_scheme:
142
+ num_bits: 8
143
+ type: float
144
+ strategy: tensor
145
+ dynamic: false
146
+ symmetric: true
147
+ """
148
+
149
+ # Apply algorithms.
150
+ oneshot(
151
+ model=model,
152
+ dataset=ds,
153
+ recipe=recipe,
154
+ max_seq_length=MAX_SEQUENCE_LENGTH,
155
+ num_calibration_samples=NUM_CALIBRATION_SAMPLES,
156
+ )
157
+
158
+ # Confirm generations of the quantized model look sane.
159
+ print("\n\n")
160
+ print("========== SAMPLE GENERATION ==============")
161
+ input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
162
+ output = model.generate(input_ids, max_new_tokens=100)
163
+ print(tokenizer.decode(output[0]))
164
+ print("==========================================\n\n")
165
+
166
+ # Save to disk compressed.
167
+ SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV"
168
+ model.save_pretrained(SAVE_DIR, save_compressed=True)
169
+ tokenizer.save_pretrained(SAVE_DIR)
170
+ ```
171
+
172
+ ## Evaluation
173
+
174
+ The model was evaluated on the [OpenLLM](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) leaderboard tasks (version 1) with the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/383bbd54bc621086e05aa1b030d8d4d5635b25e6) (commit 383bbd54bc621086e05aa1b030d8d4d5635b25e6) and the [vLLM](https://docs.vllm.ai/en/stable/) engine, using the following command:
175
+ ```
176
+ lm_eval \
177
+ --model vllm \
178
+ --model_args pretrained="neuralmagic/Phi-3.5-mini-instruct-FP8-KV",kv_cache_dtype="fp8",gpu_memory_utilization=0.4,add_bos_token=True,max_model_len=4096 \
179
+ --tasks openllm \
180
+ --batch_size auto
181
+ ```
182
+
183
+ ### Accuracy
184
+
185
+ #### Open LLM Leaderboard evaluation scores
186
+
187
+ TBD