Enable Intel GPU path and lora finetune and change examples to suppor…

…t different devices (#631)

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>
Co-authored-by: Casper <casperbh.96@gmail.com>

README.md

@@ -47,6 +47,8 @@ AutoAWQ is an easy-to-use package for 4-bit quantized models. AutoAWQ speeds up
   - Your CUDA version must be CUDA 11.8 or later.
 - AMD:
   -  Your ROCm version must be compatible with Triton.
+- Intel CPU and Intel GPU:
+  - Your torch and intel_extension_for_pytorch package version should at least 2.4 for optimized performance.
 
 ### Install from PyPi
 
@@ -60,6 +62,10 @@ There are a few ways to install AutoAWQ:
     - `INSTALL_KERNELS=1 pip install git+https://github.com/casper-hansen/AutoAWQ.git`
     - NOTE: This installs https://github.com/casper-hansen/AutoAWQ_kernels
 
+3. From main branch for Intel CPU and Intel XPU optimized performance:
+    - `pip install intel_extension_for_pytorch`
+    - `pip install git+https://github.com/casper-hansen/AutoAWQ.git`
+
 ## Usage
 
 Under examples, you can find examples of how to quantize, run inference, and benchmark AutoAWQ models.
@@ -132,6 +138,9 @@ print(f'Model is quantized and saved at "{quant_path}"')
 ```python
 from awq import AutoAWQForCausalLM
 from transformers import AutoTokenizer, TextStreamer
+from awq.utils.utils import get_best_device
+
+device = get_best_device()
 
 quant_path = "TheBloke/zephyr-7B-beta-AWQ"
 
@@ -155,7 +164,7 @@ prompt = "You're standing on the surface of the Earth. "\
 tokens = tokenizer(
     prompt_template.format(prompt=prompt), 
     return_tensors='pt'
-).input_ids.cuda()
+).input_ids.to(device)
 
 # Generate output
 generation_output = model.generate(
@@ -229,18 +238,18 @@ GPU: 2x NVIDIA GeForce RTX 4090
 ### CPU
 
 - CPU: 48 cores SPR (Intel 4th Gen Xeon CPU)
-- Command: `python examples/benchmark.py --model_path <hf_model> --batch_size 1`
+- Command: `python examples/benchmark.py --model_path <hf_model> --batch_size 1 --generator hf`
 
 | Model | Version | Batch Size | Prefill Length | Decode Length | Prefill tokens/s | Decode tokens/s | Memory |
 |-------|---------|------------|----------------|---------------|-------------------|------------------|---------------|
-| Llama 2 7B | gemm | 1 | 32 | 32 | 817.86 | 70.93 | 1.94 GB (0.00%) |
-| Llama 2 7B | gemm | 1 | 2048 | 2048 | 5279.15 | 36.83 | 2.31 GB (0.00%) |
-| Falcon | gemm | 1 | 32 | 32 | 337.51 | 26.41 | 9.57 GB (0.01%) |
-| Falcon | gemm | 1 | 2048 | 2048 | 546.71 | 18.8 | 13.46 GB (0.01%) |
-| Mistral | gemm | 1 | 32 | 32 | 343.08 | 28.46 | 9.74 GB (0.01%) |
-| Mistral | gemm | 1 | 2048 | 2048 | 1135.23 | 13.23 | 10.35 GB (0.01%) |
-| Vicuna | gemm | 1 | 32 | 32 | 340.73 | 28.86 | 9.59 GB (0.01%) |
-| Vicuna | gemm | 1 | 2048 | 2048 | 1143.19 | 11.14 | 10.98 GB (0.01%) |
+| TinyLlama 1B | gemm | 1 | 32 | 32 | 817.86 | 70.93 | 1.94 GB (0.00%) |
+| TinyLlama 1B | gemm | 1 | 2048 | 2048 | 5279.15 | 36.83 | 2.31 GB (0.00%) |
+| Falcon 7B | gemm | 1 | 32 | 32 | 337.51 | 26.41 | 9.57 GB (0.01%) |
+| Falcon 7B | gemm | 1 | 2048 | 2048 | 546.71 | 18.8 | 13.46 GB (0.01%) |
+| Mistral 7B | gemm | 1 | 32 | 32 | 343.08 | 28.46 | 9.74 GB (0.01%) |
+| Mistral 7B | gemm | 1 | 2048 | 2048 | 1135.23 | 13.23 | 10.35 GB (0.01%) |
+| Vicuna 7B | gemm | 1 | 32 | 32 | 340.73 | 28.86 | 9.59 GB (0.01%) |
+| Vicuna 7B | gemm | 1 | 2048 | 2048 | 1143.19 | 11.14 | 10.98 GB (0.01%) |
 | Llama 2 13B | gemm | 1 | 32 | 32 | 220.79 | 18.14 | 17.46 GB (0.02%) |
 | Llama 2 13B | gemm | 1 | 2048 | 2048 | 650.94 | 6.54 | 19.84 GB (0.02%) |
 | DeepSeek Coder 33B | gemm | 1 | 32 | 32 | 101.61 | 8.58 | 40.80 GB (0.04%) |

awq/models/base.py

@@ -447,7 +447,7 @@ def from_quantized(
             bool, Doc("Whether to map the weights to ExLlamaV2 kernels.")
         ] = False,
         use_ipex: Annotated[
-            bool, Doc("Whether to map the weights to ipex kernels for CPU device.")
+            bool, Doc("Whether to map the weights to ipex kernels for CPU and XPU device.")
         ] = False,
         device_map: Annotated[
             Union[str, Dict],
@@ -500,8 +500,9 @@ def from_quantized(
                 trust_remote_code=trust_remote_code,
             )
 
-        use_cpu_ipex = use_ipex or get_best_device() == "cpu"
-        if use_cpu_ipex and not ipex_available:
+        best_device = get_best_device()
+        use_ipex = use_ipex or best_device in ["cpu", "xpu:0"]
+        if use_ipex and not ipex_available:
             raise ImportError(
                 "Please install intel_extension_for_pytorch with "
                 "`pip install intel_extension_for_pytorch` for 'ipex' kernel!"
@@ -514,7 +515,7 @@ def from_quantized(
             quant_config.version,
             use_exllama=use_exllama,
             use_exllama_v2=use_exllama_v2,
-            use_ipex=use_cpu_ipex,
+            use_ipex=use_ipex,
         )
 
         model.tie_weights()
@@ -534,14 +535,12 @@ def from_quantized(
         # Dispath to devices
         awq_ext, msg = try_import("awq_ext")
         if fuse_layers:
-            if awq_ext is None:
+            if best_device in ["mps", "cuda:0"] and awq_ext is None:
                 warnings.warn("Skipping fusing modules because AWQ extension is not installed." + msg)
             else:
                 self.fuse_layers(model)
 
-        if use_cpu_ipex:
-            dtype = torch.bfloat16
-            model.to(dtype=dtype, device="cpu")
+        if use_ipex:
             # repack qweight to match the ipex kernel.
             model = ipex_post_init(model)
         elif quant_config.version == "marlin":

awq/modules/fused/attn.py

@@ -29,7 +29,7 @@ def __init__(self, head_dim, max_seq_len, device, rope_theta):
         super(RoPE, self).__init__()
 
         self.freqs_cis = nn.Parameter(
-            self.precompute_freqs_cis(head_dim, max_seq_len * 2, rope_theta).to(device),
+            self.precompute_freqs_cis(head_dim, max_seq_len, rope_theta).to(device),
             requires_grad=False,
         )
 
@@ -137,8 +137,8 @@ def __init__(
         self.use_alibi = use_alibi
         self.cache_batch_size = int(os.getenv("AWQ_BATCH_SIZE", "1"))
 
-        if kwargs.get("max_new_tokens") is not None:
-            max_seq_len = kwargs["max_new_tokens"]
+        if kwargs.get("max_length") is not None:
+            max_seq_len = kwargs["max_length"]
 
         self.max_seq_len = max_seq_len
         self.is_hf_transformers = False

awq/modules/linear/gemm_ipex.py

@@ -1,18 +1,20 @@
 import torch
 import torch.nn as nn
+from .gemm import WQLinear_GEMM
+from awq.utils.packing_utils import dequantize_gemm
 
 try:
-    from intel_extension_for_pytorch.nn.modules.weight_only_quantization import WeightOnlyQuantizedLinear
-    assert hasattr(WeightOnlyQuantizedLinear, "from_weight"), "The minimum version for ipex is at least 2.4"
+    from intel_extension_for_pytorch.llm.quantization import IPEXWeightOnlyQuantizedLinear
+    assert hasattr(IPEXWeightOnlyQuantizedLinear, "from_weight"), "The minimum version for ipex is at least 2.4"
     IPEX_INSTALLED = True
 except:
     IPEX_INSTALLED = False
 
 
-class WQLinear_IPEX(nn.Module):
+class WQLinear_IPEX(WQLinear_GEMM):
 
-    def __init__(self, w_bit, group_size, in_features, out_features, bias, dev):
-        super().__init__()
+    def __init__(self, w_bit, group_size, in_features, out_features, bias, dev, training=False):
+        nn.Module.__init__(self)
         assert IPEX_INSTALLED, \
             "Please install IPEX package with `pip install intel_extension_for_pytorch`."
         assert w_bit == 4, "Only 4 bit are supported for now."
@@ -24,12 +26,15 @@ def __init__(self, w_bit, group_size, in_features, out_features, bias, dev):
         self.w_bit = w_bit
         self.group_size = group_size if group_size != -1 else in_features
         self.scale_dtype = torch.float32
+        self.training = training
 
         # quick sanity check (make sure aligment)
         assert self.in_features % self.group_size == 0
         assert out_features % (32 // self.w_bit) == 0
         self.pack_num = 32 // self.w_bit
 
+        self.init_ipex = False
+
         self.register_buffer(
             "qzeros",
             torch.zeros(
@@ -59,10 +64,13 @@ def __init__(self, w_bit, group_size, in_features, out_features, bias, dev):
         self.register_buffer("qweight", qweight)
 
     def post_init(self):
-        assert self.qweight.device.type == "cpu"
-        self.ipex_linear = WeightOnlyQuantizedLinear.from_weight(self.qweight, self.scales, self.qzeros, \
-                                                                self.in_features, self.out_features, None, self.bias, \
-                                                                self.group_size, None, 0, 1)
+        assert self.qweight.device.type in ("cpu", "xpu")
+
+    def init_ipex_linear(self):
+        if not self.training:
+            self.ipex_linear = IPEXWeightOnlyQuantizedLinear.from_weight(self.qweight, self.scales, self.qzeros, \
+                                                                    self.in_features, self.out_features, None, self.bias, \
+                                                                    self.group_size, None, quant_method=1, dtype=0)
 
     @classmethod
     def from_linear(cls, linear, w_bit, group_size, init_only=False, scales=None):
@@ -79,16 +87,31 @@ def from_linear(cls, linear, w_bit, group_size, init_only=False, scales=None):
 
         raise NotImplementedError("Only inference is supported for IPEX kernels")
 
-    @torch.no_grad()
     def forward(self, x):
         assert IPEX_INSTALLED, (
             "IPEX kernels could not be loaded. "
             "Please install with `pip install intel_extension_for_pytorch` and "
             "refer to the detial https://github.com/intel/intel-extension-for-pytorch/tree/main")
 
-        outputs = self.ipex_linear(x)
+        if not self.init_ipex:
+            self.init_ipex_linear()
+            self.init_ipex = True
+
+        if hasattr(self, "ipex_linear"):
+            with torch.no_grad():
+                outputs = self.ipex_linear(x)
+        else:
+            outputs = dequantize_gemm(self.qweight, self.qzeros, self.scales, self.w_bit, self.group_size).to(x.dtype)
+            outputs = torch.matmul(x, outputs)
 
         return outputs
+
+    def backward(self, grad_output):
+        weights = dequantize_gemm(self.qweight, self.qzeros, self.scales, self.w_bit, self.group_size).to(grad_output.dtype)
+        batch_size = grad_output.shape[0]
+        grad_input = grad_output.bmm(weights.transpose(0, 1).unsqueeze(0).repeat(batch_size, 1, 1))
+
+        return grad_input, None, None, None, None, None, None, None
 
     def extra_repr(self) -> str:
         return ("in_features={}, out_features={}, bias={}, w_bit={}, group_size={}".format(

awq/utils/fused_utils.py

@@ -67,7 +67,9 @@ def fuse_qkv(module, q_proj, k_proj, v_proj):
         else None
     )
 
-    if isinstance(q_proj, WQLinear_GEMV):
+    if isinstance(q_proj, WQLinear_IPEX):
+        q_linear = WQLinear_IPEX
+    elif isinstance(q_proj, WQLinear_GEMV):
         q_linear = WQLinear_GEMV
     elif isinstance(q_proj, WQLinear_GEMM):
         q_linear = WQLinear_GEMM
@@ -79,8 +81,6 @@ def fuse_qkv(module, q_proj, k_proj, v_proj):
         q_linear = WQLinear_Marlin
     elif isinstance(q_proj, WQLinear_GEMVFast):
         q_linear = WQLinear_GEMVFast
-    elif isinstance(q_proj, WQLinear_IPEX):
-        q_linear = WQLinear_IPEX
 
     qkv_layer = q_linear(
         q_proj.w_bit,

awq/utils/utils.py

@@ -91,6 +91,8 @@ def get_best_device():
         return "mps"
     elif torch.cuda.is_available():
         return "cuda:0"
+    elif torch.xpu.is_available():
+        return "xpu:0"
     else:
         return "cpu"
 

examples/benchmark.py

@@ -11,9 +11,9 @@
 from transformers import AutoTokenizer, GenerationConfig, LogitsProcessor, LogitsProcessorList
 
 DEVICE = get_best_device()
-if DEVICE == "cpu":
+if DEVICE in ["cpu", "xpu:0"]:
     if ipex_available:
-        torch_dtype = torch.bfloat16
+        torch_dtype = torch.bfloat16 if DEVICE == "cpu" else torch.float16
     else:
         raise ImportError("Please import intel_extension_for_pytorch "
                           "by `pip install intel_extension_for_pytorch`")
@@ -29,8 +29,10 @@ def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
         """The logit processor is called after the model forward."""
 
         # cuda runs async operates, so we synchronize for accurate time measurement
-        if DEVICE != "cpu":
+        if DEVICE == "cuda:0":
             torch.cuda.synchronize()
+        elif DEVICE == "xpu:0":
+            torch.xpu.synchronize()
 
         # measure time
         start_time = time.time()
@@ -56,8 +58,10 @@ def generate_torch(model, input_ids, n_generate):
 
     with torch.inference_mode():
         for i in range(n_generate):
-            if DEVICE != "cpu":
+            if DEVICE == "cuda:0":
                 torch.cuda.synchronize()
+            elif DEVICE == "xpu:0":
+                torch.xpu.synchronize()
             start = time.time()
 
             if i == 0:
@@ -69,8 +73,10 @@ def generate_torch(model, input_ids, n_generate):
 
             out = model(inputs, use_cache=True)
 
-            if DEVICE != "cpu":
+            if DEVICE == "cuda:0":
                 torch.cuda.synchronize()
+            elif DEVICE == "xpu:0":
+                torch.xpu.synchronize()
             token = out[0][:, -1].max(1)[1].unsqueeze(1)
 
             if i == 0:
@@ -102,7 +108,7 @@ def generate_hf(model: BaseAWQForCausalLM, input_ids, n_generate):
 
     return context_time, generate_time
 
-def run_round(generator, model_path, quant_file, n_generate, input_ids, batch_size, no_safetensors, pretrained):
+def run_round(generator, model_path, quant_file, n_generate, context, input_ids, batch_size, no_safetensors, pretrained):
     print(f" -- Loading model...")
 
     if pretrained:
@@ -114,7 +120,7 @@ def run_round(generator, model_path, quant_file, n_generate, input_ids, batch_si
         )
     else:
         model = AutoAWQForCausalLM.from_quantized(
-            model_path, quant_file, max_seq_len=n_generate, batch_size=batch_size, safetensors=not no_safetensors
+            model_path, quant_file, max_seq_len=n_generate+context, batch_size=batch_size, safetensors=not no_safetensors
         )
 
     print(f" -- Warming up...")
@@ -149,6 +155,12 @@ def run_round(generator, model_path, quant_file, n_generate, input_ids, batch_si
             memory_pct = mem_info.rss / memory_info.total
             total_memory_used = float(mem_info.rss) / (1024 ** 3)
             print(f" ** Max Memory (device: {DEVICE}): {total_memory_used:.2f} GB ({memory_pct:.2f}%)")
+        elif DEVICE == "xpu:0":
+            for device in range(torch.xpu.device_count()):
+                memory_used = torch.xpu.max_memory_allocated(device) / (1024 ** 3)
+                total_memory_used += memory_used
+                memory_pct = memory_used / (torch.xpu.get_device_properties(device).total_memory / (1024 ** 3)) * 100
+                print(f" ** Max Memory (device: {device}): {memory_used:.2f} GB ({memory_pct:.2f}%)")
         else:
             for device in range(torch.cuda.device_count()):
                 memory_used = torch.cuda.max_memory_allocated(device) / (1024 ** 3)
@@ -197,14 +209,17 @@ def main(args):
 
     for settings in rounds:
         input_ids = torch.randint(0, tokenizer.vocab_size, (args.batch_size, settings["context"]))
-        if DEVICE != "cpu":
+        if DEVICE == "cuda:0":
             input_ids = input_ids.cuda()
+        elif DEVICE == "xpu:0":
+            input_ids = input_ids.to("xpu:0")
 
         stats, model_version = run_round(
             generator,
             args.model_path,
             args.quant_file,
             settings["n_generate"],
+            settings["context"],
             input_ids,
             args.batch_size,
             args.no_safetensors,
@@ -218,8 +233,10 @@ def main(args):
 
     df = pd.DataFrame(all_stats)
     print('Device:', DEVICE)
-    if DEVICE != "cpu":
+    if DEVICE == "cuda:0":
         print('GPU:', torch.cuda.get_device_name())
+    elif DEVICE == "xpu:0":
+        print('XPU:', torch.xpu.get_device_name())
     print('Model:', args.model_path)
     print('Version:', model_version)
     print(df.to_markdown(index=False))

examples/generate.py

@@ -1,7 +1,9 @@
 import torch
 from awq import AutoAWQForCausalLM
 from transformers import AutoTokenizer, TextStreamer
+from awq.utils.utils import get_best_device
 
+device = get_best_device()
 model_id = "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4"
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
@@ -26,7 +28,7 @@
   add_generation_prompt=True,
   return_tensors="pt",
   return_dict=True,
-).to("cuda")
+).to(device)
 
 outputs = model.generate(
     **inputs,