sharpenb commited on
Commit
5dac76a
1 Parent(s): 28a632f

Upload folder using huggingface_hub (#2)

Browse files

- 273a75e43a215ee0df3430acbceac21d937b899c8fc70d6590c9a6ced67737e6 (5de50920d56f12131000c9760d5933224d12ddf0)
- a3b96359bbc48b07ee3717aacae01e7bf2bd0ca67e4d16b6ac0d1a012d72e6a9 (a75fc3ca37c94973dcd1919e0019f30c61085110)

Files changed (5) hide show
  1. README.md +6 -3
  2. config.json +2 -2
  3. model.safetensors +2 -2
  4. plots.png +0 -0
  5. smash_config.json +1 -1
README.md CHANGED
@@ -1,5 +1,4 @@
1
  ---
2
- library_name: pruna-engine
3
  thumbnail: "https://assets-global.website-files.com/646b351987a8d8ce158d1940/64ec9e96b4334c0e1ac41504_Logo%20with%20white%20text.svg"
4
  metrics:
5
  - memory_disk
@@ -8,6 +7,8 @@ metrics:
8
  - inference_throughput
9
  - inference_CO2_emissions
10
  - inference_energy_consumption
 
 
11
  ---
12
  <!-- header start -->
13
  <!-- 200823 -->
@@ -33,13 +34,14 @@ metrics:
33
 
34
  ## Results
35
 
36
- Detailed efficiency metrics coming soon!
37
 
38
  **Frequently Asked Questions**
39
  - ***How does the compression work?*** The model is compressed with llm-int8.
40
  - ***How does the model quality change?*** The quality of the model output might vary compared to the base model.
41
  - ***How is the model efficiency evaluated?*** These results were obtained on NVIDIA A100-PCIE-40GB with configuration described in `model/smash_config.json` and are obtained after a hardware warmup. The smashed model is directly compared to the original base model. Efficiency results may vary in other settings (e.g. other hardware, image size, batch size, ...). We recommend to directly run them in the use-case conditions to know if the smashed model can benefit you.
42
  - ***What is the model format?*** We use safetensors.
 
43
  - ***What is the naming convention for Pruna Huggingface models?*** We take the original model name and append "turbo", "tiny", or "green" if the smashed model has a measured inference speed, inference memory, or inference energy consumption which is less than 90% of the original base model.
44
  - ***How to compress my own models?*** You can request premium access to more compression methods and tech support for your specific use-cases [here](https://z0halsaff74.typeform.com/pruna-access?typeform-source=www.pruna.ai).
45
  - ***What are "first" metrics?*** Results mentioning "first" are obtained after the first run of the model. The first run might take more memory or be slower than the subsequent runs due cuda overheads.
@@ -59,12 +61,13 @@ You can run the smashed model with these steps:
59
  from transformers import AutoModelForCausalLM, AutoTokenizer
60
 
61
  model = AutoModelForCausalLM.from_pretrained("PrunaAI/mosaicml-mpt-7b-8k-bnb-4bit-smashed",
62
- trust_remote_code=True)
63
  tokenizer = AutoTokenizer.from_pretrained("mosaicml/mpt-7b-8k")
64
 
65
  input_ids = tokenizer("What is the color of prunes?,", return_tensors='pt').to(model.device)["input_ids"]
66
 
67
  outputs = model.generate(input_ids, max_new_tokens=216)
 
68
  ```
69
 
70
  ## Configurations
 
1
  ---
 
2
  thumbnail: "https://assets-global.website-files.com/646b351987a8d8ce158d1940/64ec9e96b4334c0e1ac41504_Logo%20with%20white%20text.svg"
3
  metrics:
4
  - memory_disk
 
7
  - inference_throughput
8
  - inference_CO2_emissions
9
  - inference_energy_consumption
10
+ tags:
11
+ - pruna-ai
12
  ---
13
  <!-- header start -->
14
  <!-- 200823 -->
 
34
 
35
  ## Results
36
 
37
+ ![image info](./plots.png)
38
 
39
  **Frequently Asked Questions**
40
  - ***How does the compression work?*** The model is compressed with llm-int8.
41
  - ***How does the model quality change?*** The quality of the model output might vary compared to the base model.
42
  - ***How is the model efficiency evaluated?*** These results were obtained on NVIDIA A100-PCIE-40GB with configuration described in `model/smash_config.json` and are obtained after a hardware warmup. The smashed model is directly compared to the original base model. Efficiency results may vary in other settings (e.g. other hardware, image size, batch size, ...). We recommend to directly run them in the use-case conditions to know if the smashed model can benefit you.
43
  - ***What is the model format?*** We use safetensors.
44
+ - ***What calibration data has been used?*** If needed by the compression method, we used WikiText as the calibration data.
45
  - ***What is the naming convention for Pruna Huggingface models?*** We take the original model name and append "turbo", "tiny", or "green" if the smashed model has a measured inference speed, inference memory, or inference energy consumption which is less than 90% of the original base model.
46
  - ***How to compress my own models?*** You can request premium access to more compression methods and tech support for your specific use-cases [here](https://z0halsaff74.typeform.com/pruna-access?typeform-source=www.pruna.ai).
47
  - ***What are "first" metrics?*** Results mentioning "first" are obtained after the first run of the model. The first run might take more memory or be slower than the subsequent runs due cuda overheads.
 
61
  from transformers import AutoModelForCausalLM, AutoTokenizer
62
 
63
  model = AutoModelForCausalLM.from_pretrained("PrunaAI/mosaicml-mpt-7b-8k-bnb-4bit-smashed",
64
+ trust_remote_code=True, device_map='auto')
65
  tokenizer = AutoTokenizer.from_pretrained("mosaicml/mpt-7b-8k")
66
 
67
  input_ids = tokenizer("What is the color of prunes?,", return_tensors='pt').to(model.device)["input_ids"]
68
 
69
  outputs = model.generate(input_ids, max_new_tokens=216)
70
+ tokenizer.decode(outputs[0])
71
  ```
72
 
73
  ## Configurations
config.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "_name_or_path": "/tmp/tmpk4rtl8hs",
3
  "architectures": [
4
  "MPTForCausalLM"
5
  ],
@@ -65,7 +65,7 @@
65
  "quantization_config": {
66
  "bnb_4bit_compute_dtype": "bfloat16",
67
  "bnb_4bit_quant_type": "fp4",
68
- "bnb_4bit_use_double_quant": true,
69
  "llm_int8_enable_fp32_cpu_offload": false,
70
  "llm_int8_has_fp16_weight": false,
71
  "llm_int8_skip_modules": [
 
1
  {
2
+ "_name_or_path": "/tmp/tmplndb8c2a",
3
  "architectures": [
4
  "MPTForCausalLM"
5
  ],
 
65
  "quantization_config": {
66
  "bnb_4bit_compute_dtype": "bfloat16",
67
  "bnb_4bit_quant_type": "fp4",
68
+ "bnb_4bit_use_double_quant": false,
69
  "llm_int8_enable_fp32_cpu_offload": false,
70
  "llm_int8_has_fp16_weight": false,
71
  "llm_int8_skip_modules": [
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f9be3bf8f32b28b999be80082dece0c6d0eb444f4be343849d92a35e89ac7f6d
3
- size 3737395225
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56cc8af1155538787da51eb0aa49dee9b3d1e80fa74e9116d25679880436758d
3
+ size 4037639312
plots.png ADDED
smash_config.json CHANGED
@@ -8,7 +8,7 @@
8
  "compilers": "None",
9
  "task": "text_text_generation",
10
  "device": "cuda",
11
- "cache_dir": "/ceph/hdd/staff/charpent/.cache/modelsaxohfuzx",
12
  "batch_size": 1,
13
  "model_name": "mosaicml/mpt-7b-8k",
14
  "pruning_ratio": 0.0,
 
8
  "compilers": "None",
9
  "task": "text_text_generation",
10
  "device": "cuda",
11
+ "cache_dir": "/ceph/hdd/staff/charpent/.cache/modelsvwl4nos9",
12
  "batch_size": 1,
13
  "model_name": "mosaicml/mpt-7b-8k",
14
  "pruning_ratio": 0.0,