Optimize flash bert path for hpu device (#509)

Signed-off-by: kaixuanliu <kaixuan.liu@intel.com>

Author: kaixuanliu

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

@@ -2,13 +2,13 @@
 from pathlib import Path
 from torch import nn
 import torch.nn.functional as F
-from typing import Type, List
+from typing import Type, List, Union
 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.models.types import FlashBatch, Embedding, PaddedBatch
 from text_embeddings_server.utils.flash_attn import attention
 from text_embeddings_server.utils.device import use_ipex
 
@@ -166,22 +166,41 @@ def __init__(self, prefix, handle, device, dtype, config: BertConfig):
         self.num_heads = config.num_attention_heads
         self.device = device
 
-    def forward(self, hidden_states, cu_seqlens, max_s):
+    def forward(self, hidden_states, cu_seqlens, max_s, attn_mask=None):
         residual = hidden_states
-
-        qkv = torch.addmm(self.qkv_bias, hidden_states, self.qkv_weight)
-        q, k, v = qkv.view(-1, self.num_heads * 3, self.head_size).split(
-            self.num_heads, dim=1
-        )
-
+        qkv = F.linear(hidden_states, self.qkv_weight.T, self.qkv_bias)
+        bs = 1
+        hidden_dim = hidden_states.size(-1)
+        is_flat = True
+        if hidden_states.dim() > 2:
+            is_flat = False
+            bs = hidden_states.size(0)
+            q, k, v = qkv.view(bs, -1, self.num_heads * 3, self.head_size).split(
+                self.num_heads, dim=2
+            )
+        else:
+            q, k, v = qkv.view(-1, self.num_heads * 3, self.head_size).split(
+                self.num_heads, dim=1
+            )
         attn_output = torch.empty_like(q)
-        attention(q, k, v, attn_output, cu_seqlens, max_s, self.softmax_scale)
+        attention(
+            q,
+            k,
+            v,
+            attn_output,
+            cu_seqlens,
+            max_s,
+            self.softmax_scale,
+            attn_mask=attn_mask,
+        )
 
         hidden_states = torch.addmm(
             self.dense_bias,
             attn_output.view(-1, self.num_heads * self.head_size),
             self.dense_weight,
         )
+        if not is_flat:
+            hidden_states = hidden_states.view(bs, -1, hidden_dim)
         hidden_states, _ = self.layer_norm.forward(hidden_states, residual)
 
         return hidden_states
@@ -224,19 +243,16 @@ def __init__(self, prefix, handle, device, dtype, config: BertConfig):
             f"{prefix}.output.LayerNorm", handle, device, dtype, config
         )
 
-    def forward(self, hidden_states, cu_seqlens, max_s):
-        hidden_states = self.attention.forward(hidden_states, cu_seqlens, max_s)
+    def forward(self, hidden_states, cu_seqlens, max_s, attn_mask=None):
+        hidden_states = self.attention.forward(
+            hidden_states, cu_seqlens, max_s, attn_mask
+        )
         residual = hidden_states
-
-        hidden_states = torch.addmm(
-            self.intermediate_bias, hidden_states, self.intermediate_weight
+        hidden_states = F.linear(
+            hidden_states, self.intermediate_weight.T, self.intermediate_bias
         )
         hidden_states = self.intermediate_act_fn(hidden_states)
-        hidden_states = torch.addmm(
-            self.output_bias,
-            hidden_states,
-            self.output_weight,
-        )
+        hidden_states = F.linear(hidden_states, self.output_weight.T, self.output_bias)
         hidden_states, _ = self.layer_norm.forward(hidden_states, residual)
         return hidden_states
 
@@ -248,9 +264,9 @@ def __init__(self, prefix, handle, device, dtype, config: BertConfig):
             for i in range(config.num_hidden_layers)
         ]
 
-    def forward(self, hidden_states, cu_seqlens, max_s):
+    def forward(self, hidden_states, cu_seqlens, max_s, attn_mask=None):
         for layer in self.layers:
-            hidden_states = layer.forward(hidden_states, cu_seqlens, max_s)
+            hidden_states = layer.forward(hidden_states, cu_seqlens, max_s, attn_mask)
         return hidden_states
 
 
@@ -259,10 +275,21 @@ def __init__(self, handle, device, dtype, config: BertConfig):
         self.embeddings = BertEmbeddings("embeddings", handle, device, dtype, config)
         self.encoder = BertEncoder("encoder", handle, device, dtype, config)
 
-    def forward(self, input_ids, token_type_ids, position_ids, cu_seqlens, max_s):
+    def forward(
+        self,
+        input_ids,
+        token_type_ids,
+        position_ids,
+        cu_seqlens,
+        max_s,
+        mask=None,
+        attn_mask=None,
+    ):
         embeddings = self.embeddings.forward(input_ids, token_type_ids, position_ids)
-        encoder_outputs = self.encoder.forward(embeddings, cu_seqlens, max_s)
-
+        encoder_outputs = self.encoder.forward(embeddings, cu_seqlens, max_s, attn_mask)
+        if mask is not None:
+            outputs = encoder_outputs[mask]
+            return outputs[cu_seqlens[:-1]]
         return encoder_outputs[cu_seqlens[:-1]]
 
 
@@ -277,6 +304,7 @@ def __init__(self, model_path: Path, device: torch.device, dtype: torch.dtype):
 
         with safe_open(model_path / "model.safetensors", framework="pt") as f:
             model = FlashBertModel(f, device, dtype, config)
+        self.device = device
         if device.type == "hpu":
             from habana_frameworks.torch.hpu import wrap_in_hpu_graph
 
@@ -286,17 +314,38 @@ def __init__(self, model_path: Path, device: torch.device, dtype: torch.dtype):
         super(FlashBert, self).__init__(model=model, dtype=dtype, device=device)
 
     @property
-    def batch_type(self) -> Type[FlashBatch]:
-        return FlashBatch
+    def batch_type(self) -> Union[FlashBatch, PaddedBatch]:
+        # for hpu devices, we use PaddedBatch as we do not have real varlen fwd yet
+        return FlashBatch if self.device.type != "hpu" else PaddedBatch
 
     @tracer.start_as_current_span("embed")
-    def embed(self, batch: FlashBatch) -> List[Embedding]:
+    def embed(self, batch: Union[FlashBatch, PaddedBatch]) -> List[Embedding]:
+        if isinstance(batch, PaddedBatch):
+            input_lens = batch.attention_mask.cumsum(-1)[:, -1].to(torch.int32)
+            max_input_lens = input_lens.max().item()
+            cu_seqlens = torch.cat(
+                (input_lens.new_tensor([0]), input_lens.cumsum(-1).int())
+            )
+            mask = batch.attention_mask.to(torch.bool)
+            batch_size = input_lens.size(0)
+            attn_mask = torch.empty(
+                [batch_size, 1, 1, mask.shape[-1]], device=self.device
+            ).fill_(float("-inf"))
+            attn_mask[:, :, :, :].masked_fill_(mask[:, None, None, :], 0)
+        elif isinstance(batch, FlashBatch):
+            cu_seqlens = batch.cu_seqlens
+            mask = None
+            attn_mask = None
+            max_input_lens = batch.max_s
+
         embedding = self.model.forward(
             input_ids=batch.input_ids,
             token_type_ids=batch.token_type_ids,
             position_ids=batch.position_ids,
-            cu_seqlens=batch.cu_seqlens,
-            max_s=batch.max_s,
+            cu_seqlens=cu_seqlens,
+            max_s=max_input_lens,
+            mask=mask,
+            attn_mask=attn_mask,
         )
         cpu_results = embedding.view(-1).tolist()
 

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

@@ -6,6 +6,13 @@
 import torch
 import subprocess
 
+ALLOW_REDUCED_PRECISION = os.getenv(
+    "ALLOW_REDUCED_PRECISION_FP16_BF16", "true"
+).lower() in [
+    "true",
+    "1",
+]
+
 
 def _is_ipex_available():
     def get_major_and_minor_from_version(full_version):
@@ -55,6 +62,9 @@ def get_device():
     elif is_hpu():
         import habana_frameworks.torch.core as htcore
 
+        # WA for perf degradation from pytorch 2.5
+        if ALLOW_REDUCED_PRECISION:
+            torch._C._set_math_sdp_allow_fp16_bf16_reduction(True)
         if hasattr(torch, "hpu") and torch.hpu.is_available():  # type: ignore
             device = torch.device("hpu")
     elif use_ipex():

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

@@ -62,6 +62,7 @@ def hpu_attn(
     k,
     v,
     out,
+    attn_mask,
     seqlen_q,
     seqlen_k,
     max_seqlen_q,
@@ -71,66 +72,21 @@ def hpu_attn(
 ):
     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)
+    q = q.transpose(1, 2)
+    k = k.transpose(1, 2)
+    v = v.transpose(1, 2)
     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])
+    out_ = FusedSDPA.apply(q, k, v, attn_mask, 0.0, is_causal, softmax_scale)
+    out_ = out_.transpose(1, 2)
+    out.copy_(out_)
     return out
 
 
-def attention(q, k, v, out, cu_seqlens, max_s, softmax_scale, is_causal=False):
+def attention(
+    q, k, v, out, cu_seqlens, max_s, softmax_scale, is_causal=False, attn_mask=None
+):
     if HAS_FLASH_ATTN_V2:
         if use_ipex:
             import intel_extension_for_pytorch as ipex
@@ -157,6 +113,7 @@ def attention(q, k, v, out, cu_seqlens, max_s, softmax_scale, is_causal=False):
                 k,
                 v,
                 out,
+                attn_mask,
                 cu_seqlens,
                 cu_seqlens,
                 max_s,