Merge branch 'main' into hengguo/h2o

intel · Jul 29, 2024 · b2d5768 · b2d5768
2 parents 2706d5d + b00652d
commit b2d5768
Show file tree

Hide file tree

Showing 8 changed files with 12 additions and 8 deletions.
diff --git a/docs/publication.md b/docs/publication.md
@@ -1,7 +1,9 @@
 Full Publications/Events (50)
 ==========
 
-## 2024 (12)
+## 2024 (14)
+* Blog published on Intel information: [解决方案为最新Meta Llama 3.1模型提供加速](https://mp.weixin.qq.com/s/Qk3mSWPW8qdIMswpEJxVoA)
+* Blog published on Intel Developer News: [Intel AI Solutions Boost LLMs: Unleashing the Power of Meta* Llama 3.1](https://www.intel.com/content/www/us/en/developer/articles/technical/intel-ai-solutions-support-meta-llama-3-1-launch.html)
 * Blog published on digit.in: [AI hallucination in LLM and beyond: Will it ever be fixed?](https://www.digit.in/features/general/ai-hallucination-in-llm-and-beyond-can-it-be-fixed-completely.html)
 * Blog published on Medium: [Accelerating Qwen2 Models with Intel Extension for Transformers](https://medium.com/intel-analytics-software/accelerating-qwen2-models-with-intel-extension-for-transformers-99403de82f68) (June 2024)
 * Blog published on Huggingface: [Building Cost-Efficient Enterprise RAG applications with Intel Gaudi 2 and Intel Xeon](https://huggingface.co/blog/cost-efficient-rag-applications-with-intel) (May 2024)

diff --git a/examples/huggingface/pytorch/code-generation/quantization/requirements.txt b/examples/huggingface/pytorch/code-generation/quantization/requirements.txt
@@ -11,5 +11,5 @@ tiktoken #code_gen
 neural-compressor
 intel_extension_for_pytorch==2.3.0
 git+https://github.com/huggingface/optimum-intel.git@50d867c13b22c22eda451ddb67bddb8159670f85
-auto-round==0.2
+git+https://github.com/intel/auto-round.git@61cf9eef4a3ccb5a2d83a557deb709091a548581
 git+https://github.com/bigcode-project/bigcode-evaluation-harness@094c7cc197d13a53c19303865e2056f1c7488ac1
diff --git a/examples/huggingface/pytorch/text-generation/quantization/requirements_GPU.txt b/examples/huggingface/pytorch/text-generation/quantization/requirements_GPU.txt
@@ -12,6 +12,6 @@ bitsandbytes  #baichuan
 transformers_stream_generator
 tiktoken  #qwen
 einops  #qwen
-auto-round
+git+https://github.com/intel/auto-round.git@e24b9074af6cdb099e31c92eb81b7f5e9a4a244e
 git+https://github.com/intel/neural-compressor.git
 lm-eval==0.4.3
diff --git a/examples/huggingface/pytorch/text-generation/quantization/requirements_cpu_woq.txt b/examples/huggingface/pytorch/text-generation/quantization/requirements_cpu_woq.txt
@@ -11,7 +11,7 @@ transformers_stream_generator
 tiktoken  #qwen
 einops  #qwen
 git+https://github.com/intel/neural-speed.git
-auto-round==0.2
+git+https://github.com/intel/auto-round.git@e24b9074af6cdb099e31c92eb81b7f5e9a4a244e
 git+https://github.com/intel/neural-compressor.git
 lm-eval==0.4.3
 huggingface_hub
diff --git a/intel_extension_for_transformers/transformers/llm/quantization/utils.py b/intel_extension_for_transformers/transformers/llm/quantization/utils.py
@@ -658,7 +658,7 @@ def convert_to_quantized_model(model, config, device="cpu"):
                 lr=config.lr,
                 minmax_lr=config.minmax_lr,
                 seqlen=config.seq_len,
-                n_samples=config.n_samples,
+                nsamples=config.n_samples,
                 iters=config.iters,
                 scale_dtype=config.scale_dtype,
             )
@@ -672,7 +672,7 @@ def convert_to_quantized_model(model, config, device="cpu"):
                                                   dataset_name="NeelNanda/pile-10k",
                                                   seed=42,
                                                   bs=config.batch_size,
-                                                  n_samples=config.n_samples)
+                                                  nsamples=config.n_samples)
             run_fn = run_fn_for_autoround
             run_args = (dataloader,)
             model = prepare(model=model, quant_config=quant_config)

diff --git a/tests/CI/test_quantization.py b/tests/CI/test_quantization.py
@@ -432,7 +432,7 @@ def test_quantization_for_llm(self):
         woq_model.eval()
         output = woq_model(dummy_input)
         if CpuInfo().bf16:
-            self.assertTrue(isclose(float(output[0][0][0][0]), 0.1513671875, rel_tol=1e-04))
+            self.assertTrue(isclose(float(output[0][0][0][0]), 0.150390625, rel_tol=1e-04))
 
     def test_export(self):
         # test model with model_id

diff --git a/tests/CI/test_weight_only.py b/tests/CI/test_weight_only.py
@@ -208,6 +208,7 @@ def test_auto_model_saving_loading(self):
                 module_list.append(name)
         self.assertTrue(len(module_list) > 0)
 
+    @unittest.skip("need bug fix.")
     def test_nf4_training(self):
         quantization_config = RtnConfig(bits=4, weight_dtype="nf4", scale_dtype="fp32")
         model = AutoModelForCausalLM.from_pretrained(
@@ -251,6 +252,7 @@ def test_nf4_training(self):
                 module.unmerge()
         model.merge_and_unload()
 
+    @unittest.skip("need bug fix.")
     def test_int8_training(self):
         model = AutoModelForCausalLM.from_pretrained(
             llama_model_path, load_in_8bit=True, use_neural_speed=False)

diff --git a/tests/requirements.txt b/tests/requirements.txt
@@ -1,13 +1,13 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 accelerate
 auto-gptq
-auto-round==0.2
 bitsandbytes
 datasets==2.16.1
 einops
 evaluate
 gguf
 git+https://github.com/huggingface/optimum-intel.git@50d867c13b22c22eda451ddb67bddb8159670f85
+git+https://github.com/intel/auto-round.git@61cf9eef4a3ccb5a2d83a557deb709091a548581
 git+https://github.com/intel/neural-compressor.git
 git+https://github.com/intel/neural-speed.git
 intel-extension-for-pytorch==2.3.0