replace with sym quantization as there is a large accuracy drop due to kernel issue

README.md

@@ -1,21 +1,6 @@
----
-license: apache-2.0
-datasets:
-- NeelNanda/pile-10k
----
-
-
-
-
-
-
 ## Model Details
 
-This model is an int4 model with group_size 128 of [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)  generated by [intel/auto-round](https://github.com/intel/auto-round). 
-
-## How To Use
-
-
+This model is an int4 model with group_size128 and sym quantization of [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)  generated by [intel/auto-round](https://github.com/intel/auto-round).  We found there is a large accuracy drop of asym kernel for this model.
 
 
 
@@ -23,7 +8,7 @@ This model is an int4 model with group_size 128 of [microsoft/phi-2](https://hug
 
 ### INT4 Inference with AutoGPTQ
 
-Install [AutoGPTQ](https://github.com/AutoGPTQ/AutoGPTQ) from source first
+pip install auto-gptq
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -35,32 +20,44 @@ inputs = tokenizer(text, return_tensors="pt", return_attention_mask=False).to(mo
 outputs = model.generate(**inputs, max_new_tokens=50)
 text = tokenizer.batch_decode(outputs)[0]
 print(text)
+"""
+There is a girl who likes adventure,
+She loves to explore and to venture.
+She travels to faraway lands,
+And meets people from different lands.
+She learns new languages and cultures,
+And makes friends with all kinds of people.
+She is curious and brave and
+"""
 ```
 
 
 
 ### Evaluate the model 
 
-Install [lm-eval-harness 0.4.2](https://github.com/EleutherAI/lm-evaluation-harness.git) from source.
+pip install lm-eval==0.4.2
 
-Since we encountered an issue evaluating this model with lm-eval, we opted to evaluate the qdq model instead. In our assessment, we found that its accuracy closely matches that of the real quantized model in most cases except for some small models like opt-125m.
+~~bash
+lm_eval --model hf --model_args pretrained="Intel/phi-2-int4-inc" --device cuda:0 --tasks lambada_openai,hellaswag,piqa,winogrande,truthfulqa_mc1,openbookqa,boolq,arc_easy,arc_challenge,mmlu --batch_size 16
+~~
 
 
 
-| Metric         | FP16   | INT4 qdq   |
+| Metric         | FP16   | INT4   |
 | -------------- | ------ | ------ |
-| Avg.           | 0.6138 | 0.6115 |
-| mmlu           | 0.5325 | 0.5417 |
-| lambada_openai | 0.6276 | 0.6225 |
-| hellaswag      | 0.5584 | 0.5498 |
-| winogrande     | 0.7561 | 0.7545 |
-| piqa           | 0.7867 | 0.7824 |
-| truthfulqa_mc1 | 0.3146 | 0.3060 |
-| openbookqa     | 0.4020 | 0.4100 |
-| boolq          | 0.8330 | 0.8327 |
-| arc_easy       | 0.7992 | 0.7955 |
-| arc_challenge  | 0.5282 | 0.5196 |
-
+| Avg.           | 0.6131 | 0.6062 |
+| mmlu           | 0.5334 | 0.5241 |
+| lambada_openai | 0.6243 | 0.6039 |
+| hellaswag      | 0.5581 | 0.5487 |
+| winogrande     | 0.7522 | 0.7585 |
+| piqa           | 0.7867 | 0.7840 |
+| truthfulqa_mc1 | 0.3097 | 0.2974 |
+| openbookqa     | 0.4040 | 0.3960 |
+| boolq          | 0.8346 | 0.8346 |
+| arc_easy       | 0.8001 | 0.8013 |
+| arc_challenge  | 0.5282 | 0.5137 |
+
+## 
 
 ### Reproduce the model
 
@@ -76,16 +73,15 @@ python3 main.py \
 --group_size 128 \
 --bits 4 \
 --iters 1000 \
---enable_minmax_tuning \
---disable_quanted_input \
 --deployment_device 'gpu' \
---scale_dtype 'fp32' \
---eval_bs 32 \
+--disable_low_gpu_mem_usage \
 --output_dir "./tmp_autoround" \
---amp 
+
 
 ```
 
+
+
 ## Ethical Considerations and Limitations
 
 The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
@@ -106,13 +102,3 @@ Here are a couple of useful links to learn more about Intel's AI software:
 The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.
 
 
-## Cite
-
-@article{cheng2023optimize,
-  title={Optimize weight rounding via signed gradient descent for the quantization of llms},
-  author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao},
-  journal={arXiv preprint arXiv:2309.05516},
-  year={2023}
-}
-
-[arxiv](https://arxiv.org/abs/2309.05516) [github](https://github.com/intel/auto-round)

config.json

@@ -1,34 +1,31 @@
 {
   "_name_or_path": "/models/phi-2",
-  "activation_function": "gelu_new",
   "architectures": [
     "PhiForCausalLM"
   ],
-  "attn_pdrop": 0.0,
-  "auto_map": {
-    "AutoConfig": "configuration_phi.PhiConfig",
-    "AutoModelForCausalLM": "modeling_phi.PhiForCausalLM"
-  },
+  "attention_dropout": 0.0,
+  "bos_token_id": 50256,
   "embd_pdrop": 0.0,
-  "flash_attn": false,
-  "flash_rotary": false,
-  "fused_dense": false,
-  "img_processor": null,
+  "eos_token_id": 50256,
+  "hidden_act": "gelu_new",
+  "hidden_size": 2560,
   "initializer_range": 0.02,
-  "layer_norm_epsilon": 1e-05,
-  "model_type": "phi-msft",
-  "n_embd": 2560,
-  "n_head": 32,
-  "n_head_kv": null,
-  "n_inner": null,
-  "n_layer": 32,
-  "n_positions": 2048,
+  "intermediate_size": 10240,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 2048,
+  "model_type": "phi",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "partial_rotary_factor": 0.4,
+  "qk_layernorm": false,
   "quantization_config": {
-    "autoround_version": "0.1",
+    "autoround_version": "0.2.0.dev",
     "bits": 4,
     "damp_percent": 0.01,
     "desc_act": false,
     "enable_minmax_tuning": true,
+    "enable_quanted_input": true,
     "group_size": 128,
     "is_marlin_format": false,
     "iters": 1000,
@@ -37,15 +34,17 @@
     "model_file_base_name": "model",
     "model_name_or_path": null,
     "quant_method": "gptq",
+    "scale_dtype": "float16",
     "static_groups": false,
-    "sym": false,
-    "true_sequential": false,
-    "use_quant_input": false
+    "sym": true,
+    "true_sequential": false
   },
   "resid_pdrop": 0.1,
-  "rotary_dim": 32,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
   "tie_word_embeddings": false,
   "torch_dtype": "float16",
-  "transformers_version": "4.37.2",
+  "transformers_version": "4.40.2",
+  "use_cache": true,
   "vocab_size": 51200
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5ac1581429bd89dcbaf12f7dd2aeef327c71b3e37b1b15f9bd9343b63c16c3a6
-size 1836015136
+oid sha256:14c3ff2501ea2449bd14376ffe6ed545a3e6b04c0e273686acfc6f9cbe14cf22
+size 1836707656

quantize_config.json

@@ -4,16 +4,17 @@
   "damp_percent": 0.01,
   "desc_act": false,
   "static_groups": false,
-  "sym": false,
+  "sym": true,
   "true_sequential": false,
   "model_name_or_path": null,
   "model_file_base_name": "model",
   "is_marlin_format": false,
   "quant_method": "intel/auto-round",
-  "autoround_version": "0.1",
+  "autoround_version": "0.2.0.dev",
   "iters": 1000,
   "lr": 0.001,
   "minmax_lr": 0.001,
   "enable_minmax_tuning": true,
-  "use_quant_input": false
+  "enable_quanted_input": true,
+  "scale_dtype": "float16"
 }

tokenizer.json

@@ -382,6 +382,7 @@
     "end_of_word_suffix": "",
     "fuse_unk": false,
     "byte_fallback": false,
+    "ignore_merges": false,
     "vocab": {
       "!": 0,
       "\"": 1,

tokenizer_config.json

@@ -318,6 +318,7 @@
   "clean_up_tokenization_spaces": true,
   "eos_token": "<|endoftext|>",
   "model_max_length": 2048,
+  "return_token_type_ids": false,
   "tokenizer_class": "CodeGenTokenizer",
   "unk_token": "<|endoftext|>"
 }