102dee699b8136168cb1a32c8fd354718c6282d2c88e03d77ccf963ed5021857

README.md

@@ -1,5 +1,4 @@
 ---
-library_name: pruna-engine
 thumbnail: "https://assets-global.website-files.com/646b351987a8d8ce158d1940/64ec9e96b4334c0e1ac41504_Logo%20with%20white%20text.svg"
 metrics:
 - memory_disk
@@ -8,6 +7,8 @@ metrics:
 - inference_throughput
 - inference_CO2_emissions
 - inference_energy_consumption
+tags:
+- pruna-ai
 ---
 <!-- header start -->
 <!-- 200823 -->
@@ -33,7 +34,7 @@ metrics:
 
 ## Results
 
-Detailed efficiency metrics coming soon!
+![image info](./plots.png)
 
 **Frequently Asked Questions**
 - ***How does the compression work?*** The model is compressed with llm-int8.
@@ -60,7 +61,7 @@ You can run the smashed model with these steps:
    from transformers import AutoModelForCausalLM, AutoTokenizer
 
     model = AutoModelForCausalLM.from_pretrained("PrunaAI/hfl-chinese-llama-2-1.3b-bnb-4bit-smashed",
-                                                 trust_remote_code=True)
+                                                 trust_remote_code=True, device_map='auto')
     tokenizer = AutoTokenizer.from_pretrained("hfl/chinese-llama-2-1.3b")
     
     input_ids = tokenizer("What is the color of prunes?,", return_tensors='pt').to(model.device)["input_ids"]

config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "/tmp/tmpli6m8zp4",
+  "_name_or_path": "/tmp/tmpnh9hlex6",
   "architectures": [
     "LlamaForCausalLM"
   ],
@@ -22,7 +22,7 @@
   "quantization_config": {
     "bnb_4bit_compute_dtype": "bfloat16",
     "bnb_4bit_quant_type": "fp4",
-    "bnb_4bit_use_double_quant": true,
+    "bnb_4bit_use_double_quant": false,
     "llm_int8_enable_fp32_cpu_offload": false,
     "llm_int8_has_fp16_weight": false,
     "llm_int8_skip_modules": [

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:87ac2d4c3309216ee2cbbef75cc4c8854f94ff55a9f685325988245d4b5cf2e0
-size 1323696105
+oid sha256:da839b608a09a30d9d1e1b14a90216216b572a3370fb2229133a4d75bf809022
+size 1361406176

smash_config.json

@@ -8,7 +8,7 @@
         "compilers": "None",
         "task": "text_text_generation",
         "device": "cuda",
-        "cache_dir": "/ceph/hdd/staff/charpent/.cache/modelsdu1itavs",
+        "cache_dir": "/ceph/hdd/staff/charpent/.cache/models8gfbv0gp",
         "batch_size": 1,
         "model_name": "hfl/chinese-llama-2-1.3b",
         "pruning_ratio": 0.0,