[Gen] Add option to run generation with FT attention kernel

flash_attn/modules/mha.py

@@ -30,6 +30,11 @@ try:
 except ImportError:
     RotaryEmbedding = None
 
+try:
+    import ft_attention
+except ImportError:
+    ft_attention = None
+
 
 class FlashSelfAttention(nn.Module):
     """Implement the scaled dot product attention with softmax.
@@ -360,23 +365,32 @@ class MHA(nn.Module):
         assert self.layer_idx is not None, 'Generation requires layer_idx in the constructor'
         # Pre-allocate memory for key-values for inference.
         if self.layer_idx not in inference_params.key_value_memory_dict:
-            inference_kv_cache = torch.empty(
+            kv_cache = torch.empty(
                 inference_params.max_batch_size, inference_params.max_sequence_len, 2,
                 self.num_heads, self.head_dim, dtype=kv.dtype, device=kv.device
             )
-            inference_params.key_value_memory_dict[self.layer_idx] = inference_kv_cache
+            inference_params.key_value_memory_dict[self.layer_idx] = kv_cache
         else:
-            inference_kv_cache = inference_params.key_value_memory_dict[self.layer_idx]
+            assert not inference_params.fused_ft_kernel, 'fused_ft_kernel should not take this path'
+            kv_cache = inference_params.key_value_memory_dict[self.layer_idx]
         # Adjust key and value for inference
         batch_start = inference_params.batch_size_offset
         batch_end = batch_start + kv.shape[0]
-        assert batch_end <= inference_kv_cache.shape[0]
+        assert batch_end <= kv_cache.shape[0]
         sequence_start = inference_params.sequence_len_offset
         sequence_end = sequence_start + kv.shape[1]
-        assert sequence_end <= inference_kv_cache.shape[1]
+        assert sequence_end <= kv_cache.shape[1]
         # Copy key and values.
-        inference_kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
-        kv = inference_kv_cache[batch_start:batch_end, :sequence_end, ...]
+        kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
+        kv = kv_cache[batch_start:batch_end, :sequence_end, ...]
+        if inference_params.fused_ft_kernel:
+            # FT kernel requires different layouts for the k_cache and v_cache.
+            assert kv_cache.dtype in [torch.float16, torch.bfloat16, torch.float32]
+            packsize = 4 if kv_cache.dtype == torch.float32 else 8
+            k_cache = rearrange(kv_cache[:, :, 0], 'b s h (d packsize) -> b h d s packsize',
+                                packsize=packsize).contiguous()
+            v_cache = rearrange(kv_cache[:, :, 1], 'b s h d -> b h s d').contiguous()
+            inference_params.key_value_memory_dict[self.layer_idx] = (k_cache, v_cache)
         return kv
 
     def forward(self, x, x_kv=None, key_padding_mask=None, cu_seqlens=None, max_seqlen=None,
@@ -430,14 +444,24 @@ class MHA(nn.Module):
                 else:
                     context = torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, **kwargs)
             else:
-                if self.rotary_emb_dim > 0:
-                    qkv = self.rotary_emb(qkv, seqlen_offset=inference_params.sequence_len_offset)
-                q = qkv[:, :, 0]
-                kv = self._update_kv_cache(qkv[:, :, 1:], inference_params)
-                # If we're processing the prompt, causal=None (use self.causal).
-                # If we're decoding, then causal=False.
-                causal = None if inference_params.sequence_len_offset == 0 else False
-                context = self.inner_cross_attn(q, kv, causal=causal)
+                if (not inference_params.fused_ft_kernel) or inference_params.sequence_len_offset == 0:
+                    if self.rotary_emb_dim > 0:
+                        qkv = self.rotary_emb(qkv, seqlen_offset=inference_params.sequence_len_offset)
+                    q = qkv[:, :, 0]
+                    kv = self._update_kv_cache(qkv[:, :, 1:], inference_params)
+                    # If we're processing the prompt, causal=None (use self.causal).
+                    # If we're decoding, then causal=False.
+                    causal = None if inference_params.sequence_len_offset == 0 else False
+                    context = self.inner_cross_attn(q, kv, causal=causal)
+                else:
+                    assert ft_attention is not None
+                    context = ft_attention.single_query_attention(
+                        *rearrange(qkv, 'b 1 three h d -> b three h d').unbind(dim=1),
+                        *inference_params.key_value_memory_dict[self.layer_idx],
+                        inference_params.lengths_per_sample, inference_params.sequence_len_offset,
+                        self.rotary_emb_dim
+                    )
+                    context = rearrange(context, 'b h d -> b 1 h d')
         else:
             if not self.return_residual:
                 q = self.Wq(x)

flash_attn/utils/generation.py

 # Copyright (c) 2022, Tri Dao.
 # Adapted from https://github.com/NVIDIA/Megatron-LM/blob/0bb597b42c53355a567aba2a1357cc34b9d99ddd/megatron/text_generation/forward_step.py#L31
+from typing import Optional
+
 from dataclasses import dataclass, field
 import torch
+from torch import Tensor
 
 from einops import rearrange
 
@@ -17,9 +20,11 @@ class InferenceParams:
     sequence_len_offset: int = 0
     batch_size_offset: int = 0
     key_value_memory_dict: dict = field(default_factory=dict)
+    fused_ft_kernel: bool = False
+    lengths_per_sample: Optional[Tensor] = None
 
 
-def greedy_decode(input_ids, model, max_length):
+def greedy_decode(input_ids, model, max_length, fused_ft_kernel=True):
     """Greedy decoding. This is a very simple implementation.
     We assume that all sequences in the same batch have the same length.
     Arguments:
@@ -30,7 +35,8 @@ def greedy_decode(input_ids, model, max_length):
         scores: tuples of (batch, vocab_size)
     """
     batch_size, seqlen_og = input_ids.shape
-    inference_params = InferenceParams(max_sequence_len=max_length, max_batch_size=batch_size)
+    inference_params = InferenceParams(max_sequence_len=max_length, max_batch_size=batch_size,
+                                       fused_ft_kernel=fused_ft_kernel)
     scores = []
     with torch.inference_mode():
         logits = model(input_ids, inference_params=inference_params).logits[:, -1]
@@ -57,8 +63,9 @@ def greedy_decode(input_ids, model, max_length):
 
 class GenerationMixin:
 
-    def generate(self, input_ids, max_length, return_dict_in_generate=False, output_scores=False):
-        output = greedy_decode(input_ids, self, max_length)
+    def generate(self, input_ids, max_length, return_dict_in_generate=False, output_scores=False,
+                 **kwargs):
+        output = greedy_decode(input_ids, self, max_length, **kwargs)
         if not output_scores:
             output.scores = None
         return output if return_dict_in_generate else output.sequences

tests/models/test_gpt_generation.py

@@ -15,10 +15,11 @@ from flash_attn.utils.generation import greedy_decode
 
 
 # TODO: test with rotary embedding
+@pytest.mark.parametrize('fused_ft_kernel', [False, True])
 @pytest.mark.parametrize('optimized', [False, True])
-# @pytest.mark.parametrize('optimized', [False])
+# @pytest.mark.parametrize('optimized', [True])
 @pytest.mark.parametrize('model_name', ["gpt2"])
-def test_greedy_decode(model_name, optimized):
+def test_greedy_decode(model_name, optimized, fused_ft_kernel):
     """Check that our implementation of GPT2 generation matches the HF implementation:
     the scores in fp16 should be around the same as the HF scores in fp16, when compared to
     the HF scores in fp32.
@@ -62,6 +63,7 @@ def test_greedy_decode(model_name, optimized):
     scores = tuple(scores)
 
     out = model.generate(input_ids=input_ids, max_length=max_length,
+                         fused_ft_kernel=fused_ft_kernel,
                          return_dict_in_generate=True, output_scores=True)
 
     out_hf = model_hf.generate(input_ids=input_ids, max_length=max_length,