Merge pull request huggingface#5 from kaixuanliu/ipex

add hpu flashBert support

Author: yuanwu2017

Dockerfile-intel

@@ -77,7 +77,7 @@ RUN python -m pip install torch==2.4.0 torchvision torchaudio==2.4.0 --index-url
 RUN cd backends/python/server && \
     make install
 
-FROM vault.habana.ai/gaudi-docker/1.16.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.2:latest AS hpu
+FROM vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest AS hpu
 ENV HUGGINGFACE_HUB_CACHE=/data \
     PORT=80
 

backends/python/server/text_embeddings_server/models/__init__.py

@@ -37,6 +37,7 @@ def get_model(model_path: Path, dtype: Optional[str]) :
         raise RuntimeError(f"Unknown dtype {dtype}")
 
     device = get_device()
+    logger.info(f"backend device: {device}")
     config = AutoConfig.from_pretrained(model_path)
     if config.model_type == "bert":
         config: BertConfig
@@ -48,14 +49,12 @@ def get_model(model_path: Path, dtype: Optional[str]) :
         ):
             return FlashBert(model_path, device, datatype) # type: ignore
         if use_ipex() and device.type in ["cpu", "xpu"]:
+            import intel_extension_for_pytorch as ipex
             return FlashBert(model_path, device, datatype) # type: ignore
         if device.type == "hpu":
-            from habana_frameworks.torch.hpu import wrap_in_hpu_graph
-            from optimum.habana.transformers.modeling_utils import adapt_transformers_to_gaudi
-            adapt_transformers_to_gaudi()
-            model_handle = DefaultModel(model_path, device, datatype)
-            model_handle.model = wrap_in_hpu_graph(model_handle.model, disable_tensor_cache=True)
-            return model_handle
+            import habana_frameworks.torch.core as htcore
+            return FlashBert(model_path, device, datatype)
+
         return DefaultModel(model_path, device, datatype)
     else:
         return DefaultModel(model_path, device, datatype)

backends/python/server/text_embeddings_server/models/flash_bert.py

@@ -1,27 +1,43 @@
 import torch
-
 from pathlib import Path
 from torch import nn
+import torch.nn.functional as F
 from typing import Type, List
 from safetensors import safe_open
 from transformers.activations import ACT2FN
 from transformers.models.bert import BertConfig
 from opentelemetry import trace
-
-
 from text_embeddings_server.models import Model
 from text_embeddings_server.models.types import FlashBatch, Embedding
 from text_embeddings_server.utils.flash_attn import attention
 from text_embeddings_server.utils.device import use_ipex
+
 tracer = trace.get_tracer(__name__)
 
+def hpu_add_layer_norm(
+    add: torch.Tensor,
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    epsilon: float,
+    add_back: bool
+):
+    if add is not None:
+        added_tensor = torch.add(add, x, alpha=1.0)
+        output = F.layer_norm(added_tensor, [x.size(-1)], weight, bias, epsilon)
+        if add_back:
+            add.add_(x)
+        return output
+    else:
+        return F.layer_norm(x, [x.size(-1)], weight=weight, bias=bias, eps=epsilon)
 
 class FastLayerNorm:
     def __init__(self, prefix, handle, device, dtype, config: BertConfig):
         self.weight = handle.get_tensor(f"{prefix}.weight").to(dtype).to(device)
         self.bias = handle.get_tensor(f"{prefix}.bias").to(dtype).to(device)
         self.variance_epsilon = config.layer_norm_eps
         self.device = device
+        self.use_ipex = use_ipex()
 
     def forward(self, hidden_states, residual=None):
         # Flash attention imports
@@ -48,7 +64,7 @@ def forward(self, hidden_states, residual=None):
             )
             if res is None:
                 res = hidden_states
-        elif use_ipex():
+        elif self.use_ipex:
             import intel_extension_for_pytorch as ipex
             normed_hidden_states = ipex.llm.functional.add_layer_norm(
                 residual,
@@ -60,7 +76,16 @@ def forward(self, hidden_states, residual=None):
              )
 
             res = residual if residual is not None else hidden_states
-
+        elif self.device.type == "hpu":
+            normed_hidden_states = hpu_add_layer_norm(
+                residual,
+                hidden_states,
+                self.weight,
+                self.bias,
+                self.variance_epsilon,
+                residual is not None
+            )
+            res = residual if residual is not None else hidden_states
         return normed_hidden_states, res
 
 
@@ -242,7 +267,9 @@ def __init__(self, model_path: Path, device: torch.device, dtype: torch.dtype):
         config = BertConfig.from_pretrained(model_path)
         with safe_open(model_path / "model.safetensors", framework="pt") as f:
             model = FlashBertModel(f, device, dtype, config)
-
+        if device.type == "hpu":
+            from habana_frameworks.torch.hpu import wrap_in_hpu_graph
+            model = wrap_in_hpu_graph(model, disable_tensor_cache=False)
         self.hidden_size = config.hidden_size
 
         super(FlashBert, self).__init__(model=model, dtype=dtype, device=device)

backends/python/server/text_embeddings_server/utils/device.py

@@ -27,11 +27,11 @@ def get_major_and_minor_from_version(full_version):
         return False
     return True
 
-def _is_hpu() -> bool:
+def is_hpu() -> bool:
     is_hpu_available = True
     try:
         subprocess.run(["hl-smi"], capture_output=True, check=True)
-    except (FileNotFoundError, PermissionError, subprocess.CalledProcessError):
+    except:
         is_hpu_available = False
     return is_hpu_available
 
@@ -43,7 +43,7 @@ def get_device() :
     device = torch.device("cpu")
     if torch.cuda.is_available():
         device = torch.device("cuda")
-    elif _is_hpu():
+    elif is_hpu():
         import habana_frameworks.torch.core as htcore
         if hasattr(torch, "hpu") and torch.hpu.is_available(): # type: ignore
             device = torch.device("hpu")

backends/python/server/text_embeddings_server/utils/flash_attn.py

@@ -1,6 +1,6 @@
 import os
 import torch
-from text_embeddings_server.utils.device import use_ipex
+from text_embeddings_server.utils.device import use_ipex, is_hpu
 
 from loguru import logger
 
@@ -10,7 +10,10 @@
 HAS_FLASH_ATTN = False
 HAS_FLASH_ATTN_V2 = False
 
-if use_ipex():
+is_hpu = is_hpu()
+use_ipex = use_ipex()
+
+if use_ipex or is_hpu:
     HAS_FLASH_ATTN_V2 = True
 else:
     if not torch.cuda.is_available():
@@ -54,14 +57,77 @@
         HAS_FLASH_ATTN = True
 
 
+def hpu_attn(q, k, v, out, seqlen_q, seqlen_k, max_seqlen_q, max_seqlen_k, softmax_scale, is_causal=False):
+    from habana_frameworks.torch.hpex.kernels import FusedSDPA
+    total_q, num_head, head_size = q.size()
+    total_k, num_head_k, _ = k.size()
+    batch_size = seqlen_q.size(0) - 1
+    seqlen_q_ = seqlen_q.clone()
+    seqlen_q_[:batch_size] = seqlen_q[1:]
+    seqlen_q = (seqlen_q_ - seqlen_q)[:batch_size]
+    seqlen_k_ = seqlen_k.clone()
+    seqlen_k_[:batch_size] = seqlen_k[1:]
+    seqlen_k = (seqlen_k_ - seqlen_k)[:batch_size]
+
+    pad_q = torch.zeros(
+        [batch_size, max_seqlen_q, num_head, head_size],
+        dtype=q.dtype,
+        device=q.device,
+    )
+    pad_k = torch.zeros(
+        [batch_size, max_seqlen_k, num_head_k, head_size],
+        dtype=k.dtype,
+        device=k.device,
+    )
+    pad_v = torch.zeros(
+        [batch_size, max_seqlen_k, num_head_k, head_size],
+        dtype=v.dtype,
+        device=v.device,
+    )
+    q_mask = torch.arange(0, max_seqlen_q, device=q.device)[None, :].repeat(
+        batch_size, 1
+    )
+    q_mask = q_mask < seqlen_q[:, None].repeat(1, q_mask.size(-1))
+    k_mask = torch.arange(0, max_seqlen_k, device=k.device)[None, :].repeat(
+        batch_size, 1
+    )
+    k_mask = k_mask < seqlen_k[:, None].repeat(1, k_mask.size(-1))
+    align_mask_seqlen = max_seqlen_k
+    attn_mask = torch.empty(
+        [batch_size, 1, 1, align_mask_seqlen],
+        dtype=q.dtype,
+        device=q.device,
+    ).fill_(float("-inf"))
+    attn_mask[:, :, :, :max_seqlen_k].masked_fill_(k_mask[:, None, None, :], 0)
+
+    pad_q[q_mask] = q
+    pad_k[k_mask] = k
+    pad_v[k_mask] = v
+
+    pad_q = pad_q.permute(0, 2, 1, 3)
+    pad_k = pad_k.permute(0, 2, 1, 3)
+    pad_v = pad_v.permute(0, 2, 1, 3)
+    if is_causal:
+        attn_mask = None
+
+    out_ = FusedSDPA.apply(pad_q, pad_k, pad_v, attn_mask, 0.0, is_causal, softmax_scale)
+    out_ = out_.permute(0, 2, 1, 3)
+    out.copy_(out_[q_mask])
+    return out
+
+
 def attention(q, k, v, out, cu_seqlens, max_s, softmax_scale, is_causal=False):
     if HAS_FLASH_ATTN_V2:
-        if use_ipex():
+        if use_ipex:
             import intel_extension_for_pytorch as ipex
-            return ipex.llm.functional.varlen_attention(q, k, v, out, cu_seqlens, cu_seqlens,
-                                                        max_s, max_s, 0, softmax_scale,
-                                                        zero_tensors=False, is_causal=False,
+            return ipex.llm.functional.varlen_attention(q, k, v, out, cu_seqlens, cu_seqlens, 
+                                                        max_s, max_s, 0, softmax_scale, 
+                                                        zero_tensors=False, is_causal=False, 
                                                         return_softmax=False, gen_=None)
+        elif is_hpu:
+            return hpu_attn(q, k, v, out, cu_seqlens, cu_seqlens,
+                            max_s, max_s, softmax_scale, is_causal=False)
+
         else:
             return flash_attn_2_cuda.varlen_fwd(
                 q,