standardize attn hijack patches (#381)

* split sdp attn into its own patch

* sync xformers patch to follow shared format and be diffable

* update flash-attn patch for 70B/GQA and inference using helper from flash-attn tests

* speed up flash-attn inference

* fix patch to check position ids and don't use multipack for evals

* copy LlamaModel.forward and LlamaDecoderLayer.forward into monkeypatch

* update forwards so we only calculate cu_seqlens once

* enable eval dataloader using multipack again

* fix the patch to work properly and work with FSDP

---------

Co-authored-by: Wing Lian <[email protected]>

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -2,26 +2,63 @@
 
 # copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py
 
-from typing import Optional, Tuple
+import warnings
+from typing import List, Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.llama.modeling_llama import (
+    LlamaDecoderLayer as OriginalLlamaDecoderLayer,
+)
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
 try:
-    from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func
+    from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
+        flash_attn_kvpacked_func,
+        flash_attn_varlen_kvpacked_func,
+        flash_attn_varlen_qkvpacked_func,
+    )
 except ImportError:
+    from flash_attn.flash_attn_interface import (
+        flash_attn_unpadded_kvpacked_func as flash_attn_varlen_kvpacked_func,
+    )
     from flash_attn.flash_attn_interface import (
         flash_attn_unpadded_qkvpacked_func as flash_attn_varlen_qkvpacked_func,
     )
 
-from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
 
-from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
+    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
+        _prepare_decoder_attention_mask
+    )
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = flashattn_forward
+    if packed:
+        transformers.models.llama.modeling_llama.LlamaDecoderLayer = LlamaDecoderLayer
+        transformers.models.llama.modeling_llama.LlamaModel.forward = (
+            llama_model_forward
+        )
 
 
-def forward(
+# Disable the transformation of the attention mask in LlamaModel as the flash attention
+# requires the attention mask to be the same as the key_padding_mask
+def _prepare_decoder_attention_mask(
+    self,
+    attention_mask,
+    input_shape,
+    inputs_embeds,
+    past_key_values_length,
+):  # pylint: disable=unused-argument
+    # [bsz, seq_len]
+    return attention_mask
+
+
+def flashattn_forward(
     self,
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor] = None,
@@ -29,6 +66,8 @@ def forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     """Input shape: Batch x Time x Channel
 
@@ -37,124 +76,523 @@ def forward(
     # pylint: disable=duplicate-code
     bsz, q_len, _ = hidden_states.size()
 
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        value_states = torch.cat(value_states, dim=-1)
+
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+    query_states = query_states.view(
+        bsz, q_len, self.num_heads, self.head_dim
+    ).transpose(1, 2)
+    key_states = key_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    value_states = value_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
     # [bsz, q_len, nh, hd]
     # [bsz, nh, q_len, hd]
 
     kv_seq_len = key_states.shape[-2]
-    assert past_key_value is None, "past_key_value is not supported"
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
 
     cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
     query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )
     # [bsz, nh, t, hd]
-    assert not output_attentions, "output_attentions is not supported"
-    assert not use_cache, "use_cache is not supported"
-
-    # Flash attention codes from
-    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py
-
-    # transform the data into the format required by flash attention
-    qkv = torch.stack(
-        [query_states, key_states, value_states], dim=2
-    )  # [bsz, nh, 3, q_len, hd]
-    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
-    # We have disabled _prepare_decoder_attention_mask in LlamaModel
-    # the attention_mask should be the same as the key_padding_mask
-    key_padding_mask = attention_mask
-
-    if key_padding_mask is None:
-        qkv = rearrange(qkv, "b s ... -> (b s) ...")
-        max_s = q_len
-        cu_q_lens = torch.arange(
-            0,
-            (bsz + 1) * q_len,
-            step=q_len,
-            dtype=torch.int32,
-            device=qkv.device,
-        )
-        output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+
+    if past_key_value is not None:
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
         )
-        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
-    elif attention_mask.shape[0] == 1:
+
+    #
+    # flash-attn v2 start
+    #
+
+    if self.training:
+        # during training q,k,v always have same seqlen
+        assert key_states.shape == query_states.shape
+        is_causal = True
+    else:
+        # turn off FA causal mask after first inference autoregressive iteration
+        # only on first autoregressive step q,k,v have same seqlen
+        is_causal = past_key_value is not None
+
+    if cu_seqlens is not None and max_seqlen is not None:
         # special handling using sample packing
+        qkv = torch.stack(
+            [query_states, key_states, value_states], dim=2
+        )  # [bsz, nh, 3, q_len, hd]
+        qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
         qkv = rearrange(qkv, "b s ... -> (b s) ...")
-        cu_q_lens, max_s = get_cu_seqlens_from_pos_ids(position_ids)
-        cu_q_lens = cu_q_lens.squeeze()
 
         output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+            qkv, cu_seqlens, max_seqlen, 0.0, softmax_scale=None, causal=is_causal
         )
         output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
-    else:
-        nheads = qkv.shape[-2]
-
-        # pylint: disable=invalid-name
-        x = rearrange(qkv, "b s three h d -> b s (three h d)")
-        x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
-        x_unpad = rearrange(
-            x_unpad,
-            "nnz (three h d) -> nnz three h d",
-            three=3,
-            h=nheads,
+    elif query_states.shape == key_states.shape:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        qkv_unpad, cu_seqlens_q, max_seqlen_q, _, output_pad_fn = generate_qkv(
+            query_states,
+            key_states,
+            value_states,
+            qkvpacked=True,
+            # We have disabled _prepare_decoder_attention_mask in LlamaModel
+            # the attention_mask should be the same as the key_padding_mask
+            key_padding_mask=attention_mask,
+            query_padding_mask=attention_mask[:, -query_states.size(1) :]
+            if attention_mask is not None
+            else None,
         )
         output_unpad = flash_attn_varlen_qkvpacked_func(
-            x_unpad,
-            cu_q_lens,
-            max_s,
+            qkv_unpad,
+            cu_seqlens_q,
+            max_seqlen_q,
             0.0,
             softmax_scale=None,
-            causal=True,
+            causal=is_causal,
         )
-        output = rearrange(
-            pad_input(
-                rearrange(output_unpad, "nnz h d -> nnz (h d)"),
-                indices,
-                bsz,
-                q_len,
-            ),
-            "b s (h d) -> b s h d",
-            h=nheads,
+        output = output_pad_fn(output_unpad)
+    else:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        if attention_mask is None or attention_mask.all().item():
+            output = flash_attn_kvpacked_func(
+                query_states,
+                torch.stack([key_states, value_states], 2),
+                causal=is_causal,
+            )
+        else:
+            (  # pylint: disable=unbalanced-tuple-unpacking
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                _,
+                _,
+                output_pad_fn,
+            ) = generate_qkv(
+                query_states,
+                key_states,
+                value_states,
+                kvpacked=True,
+                key_padding_mask=attention_mask,
+                query_padding_mask=attention_mask[:, -query_states.size(1) :]
+                if attention_mask is not None
+                else None,
+            )
+            output_unpad = flash_attn_varlen_kvpacked_func(
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                0.0,
+                softmax_scale=None,
+                causal=is_causal,
+            )
+            output = output_pad_fn(output_unpad)
+
+    attn_output = output
+    if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, q_len, self.num_heads, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+    attn_output = rearrange(attn_output, "b s h d -> b s (h d)")
+
+    #
+    # flash-attn v2 end
+    #
+
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value
+
+
+# based on https://github.com/Dao-AILab/flash-attention/blob/364a5b/tests/test_flash_attn.py#L38
+def generate_qkv(
+    q,
+    k,
+    v,
+    query_padding_mask=None,
+    key_padding_mask=None,
+    kvpacked=False,
+    qkvpacked=False,
+):  # pylint: disable=invalid-name,unnecessary-lambda-assignment
+    """
+    Arguments:
+        q: (batch_size, seqlen_q, nheads, d)
+        k: (batch_size, seqlen_k, nheads_k, d)
+        v: (batch_size, seqlen_k, nheads_k, d)
+        query_padding_mask: (batch_size, seqlen), bool
+        key_padding_mask: (batch_size, seqlen), bool
+    """
+    assert not (kvpacked and qkvpacked)
+    batch_size, seqlen_q, nheads, d = q.shape
+    _, seqlen_k, nheads_k, _ = k.shape
+    assert k.shape == (batch_size, seqlen_k, nheads_k, d)
+    assert v.shape == (batch_size, seqlen_k, nheads_k, d)
+
+    if query_padding_mask is not None:
+        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(
+            q, query_padding_mask
+        )
+
+        output_pad_fn = lambda output_unpad: pad_input(  # noqa: E731
+            output_unpad, indices_q, batch_size, seqlen_q
+        )
+
+    else:
+        q_unpad = rearrange(q, "b s h d -> (b s) h d")
+        cu_seqlens_q = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_q,
+            step=seqlen_q,
+            dtype=torch.int32,
+            device=q_unpad.device,
+        )
+        max_seqlen_q = seqlen_q
+
+        output_pad_fn = lambda output_unpad: rearrange(  # noqa: E731
+            output_unpad, "(b s) h d -> b s h d", b=batch_size
+        )
+
+    if key_padding_mask is not None:
+        k_unpad, _, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
+        v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
+    else:
+        k_unpad = rearrange(k, "b s h d -> (b s) h d")
+        v_unpad = rearrange(v, "b s h d -> (b s) h d")
+        cu_seqlens_k = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_k,
+            step=seqlen_k,
+            dtype=torch.int32,
+            device=k_unpad.device,
+        )
+        max_seqlen_k = seqlen_k
+
+    if qkvpacked:
+        assert nheads == nheads_k
+        qkv_unpad = torch.stack([q_unpad, k_unpad, v_unpad], dim=1)
+        qkv = torch.stack([q, k, v], dim=2)
+        return (qkv_unpad, cu_seqlens_q, max_seqlen_q, qkv, output_pad_fn)
+
+    if kvpacked:
+        kv_unpad = torch.stack([k_unpad, v_unpad], dim=1)
+        kv = torch.stack([k, v], dim=2)
+        return (
+            q_unpad,
+            kv_unpad,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_q,
+            max_seqlen_k,
+            q,
+            kv,
+            output_pad_fn,
         )
 
     return (
-        self.o_proj(rearrange(output, "b s h d -> b s (h d)")),
-        None,
-        None,
+        q_unpad,
+        k_unpad,
+        v_unpad,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlen_q,
+        max_seqlen_k,
+        q,
+        k,
+        v,
+        output_pad_fn,
     )
 
 
-# Disable the transformation of the attention mask in LlamaModel as the flash attention
-# requires the attention mask to be the same as the key_padding_mask
-def _prepare_decoder_attention_mask(
+def llama_model_forward(
     self,
-    attention_mask,
-    input_shape,
-    inputs_embeds,
-    past_key_values_length,
-):  # pylint: disable=unused-argument
-    # [bsz, seq_len]
-    return attention_mask
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
 
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
 
-def replace_llama_attn_with_flash_attn():
-    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
-        _prepare_decoder_attention_mask
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+        )
+    if input_ids is not None:
+        batch_size, seq_length = input_ids.shape
+    elif inputs_embeds is not None:
+        batch_size, seq_length, _ = inputs_embeds.shape
+    else:
+        raise ValueError(
+            "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+        )
+
+    seq_length_with_past = seq_length
+    past_key_values_length = 0
+
+    if past_key_values is not None:
+        past_key_values_length = past_key_values[0][0].shape[2]
+        seq_length_with_past = seq_length_with_past + past_key_values_length
+
+    cu_seqlens = None
+    max_seqlen = None
+    if position_ids is None:
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        position_ids = torch.arange(
+            past_key_values_length,
+            seq_length + past_key_values_length,
+            dtype=torch.long,
+            device=device,
+        )
+        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+    else:
+        position_ids = position_ids.view(-1, seq_length).long()
+        cu_seqlens, max_seqlen = get_cu_seqlens_from_pos_ids(position_ids)
+        cu_seqlens = cu_seqlens.squeeze()
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+    # embed positions
+    if attention_mask is None:
+        attention_mask = torch.ones(
+            (batch_size, seq_length_with_past),
+            dtype=torch.bool,
+            device=inputs_embeds.device,
+        )
+    attention_mask = (
+        self._prepare_decoder_attention_mask(  # pylint: disable=protected-access
+            attention_mask,
+            (batch_size, seq_length),
+            inputs_embeds,
+            past_key_values_length,
+        )
+    )
+
+    hidden_states = inputs_embeds
+
+    if self.gradient_checkpointing and self.training:
+        if use_cache:
+            transformers.logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    next_decoder_cache = () if use_cache else None
+
+    for idx, decoder_layer in enumerate(self.layers):
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+        if self.gradient_checkpointing and self.training:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    # None for past_key_value
+                    return module(*inputs)
+
+                return custom_forward
+
+            layer_outputs = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(decoder_layer),
+                hidden_states,
+                attention_mask,
+                position_ids,
+                None,
+                output_attentions,
+                None,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = next_decoder_cache if use_cache else None
+    if not return_dict:
+        return tuple(
+            v
+            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+            if v is not None
+        )
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
     )
-    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward
+
+
+class LlamaDecoderLayer(OriginalLlamaDecoderLayer):
+    """
+    patched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens
+    """
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[torch.Tensor] = None,
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cu_seqlens (`torch.Tensor`, *optional*) cumulative sequence len when packing
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs

src/axolotl/monkeypatch/llama_attn_hijack_sdp.py

@@ -0,0 +1,140 @@
+"""
+Patched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention
+"""
+
+import warnings
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+import transformers.models.llama.modeling_llama
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+
+
+def hijack_llama_sdp_attention():
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
+        sdp_attention_forward
+    )
+
+
+def sdp_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    # pylint: disable=duplicate-code
+    bsz, q_len, _ = hidden_states.size()
+
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        value_states = torch.cat(value_states, dim=-1)
+
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+    query_states = query_states.view(
+        bsz, q_len, self.num_heads, self.head_dim
+    ).transpose(1, 2)
+    key_states = key_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    value_states = value_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
+    )
+    # [bsz, nh, t, hd]
+
+    if past_key_value is not None:
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    #
+    # sdp-attn start
+    #
+
+    with torch.backends.cuda.sdp_kernel():
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            is_causal=False,
+        )
+
+    if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+    attn_output = attn_output.transpose(1, 2)
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    #
+    # sdp-attn end
+    #
+
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -3,13 +3,13 @@ Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-g
 """
 
 import logging
-import math
+import warnings
 from typing import Optional, Tuple
 
 import torch
 import torch.nn.functional as F
 import transformers.models.llama.modeling_llama
-from torch import nn
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
 
 try:
     import xformers.ops
@@ -21,12 +21,6 @@ def hijack_llama_attention():
     transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
 
 
-def hijack_llama_sdp_attention():
-    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
-        sdp_attention_forward
-    )
-
-
 def xformers_forward(
     self,
     hidden_states: torch.Tensor,
@@ -81,15 +75,15 @@ def xformers_forward(
     value_states = value_states.view(
         bsz, q_len, self.num_key_value_heads, self.head_dim
     ).transpose(1, 2)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
 
     kv_seq_len = key_states.shape[-2]
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
+
     cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    (
-        query_states,
-        key_states,
-    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
+    query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )
     # [bsz, nh, t, hd]
@@ -102,74 +96,50 @@ def xformers_forward(
     past_key_value = (key_states, value_states) if use_cache else None
 
     # repeat k/v heads if n_kv_heads < n_heads
-    key_states = transformers.models.llama.modeling_llama.repeat_kv(
-        key_states, self.num_key_value_groups
-    )
-    value_states = transformers.models.llama.modeling_llama.repeat_kv(
-        value_states, self.num_key_value_groups
-    )
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
 
-    # We only apply xformers optimizations if we don't need to output the whole attention matrix
-    if not output_attentions:
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-
-        # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
-        # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
-        if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
-            # input and output should be of form (bsz, q_len, num_heads, head_dim)
-            attn_output = xformers.ops.memory_efficient_attention(
-                query_states, key_states, value_states, attn_bias=None
-            )
-        else:
-            # input and output should be of form (bsz, q_len, num_heads, head_dim)
-            attn_output = xformers.ops.memory_efficient_attention(
-                query_states,
-                key_states,
-                value_states,
-                # attn_bias=attention_mask,
-                attn_bias=xformers.ops.LowerTriangularMask(),
-            )
-        attn_weights = None
-    else:
-        attn_weights = torch.matmul(
-            query_states, key_states.transpose(2, 3)
-        ) / math.sqrt(self.head_dim)
-
-        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
-            raise ValueError(
-                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
-                f" {attn_weights.size()}"
-            )
-
-        if attention_mask is not None:
-            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
-                raise ValueError(
-                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
-                )
-            attn_weights = attn_weights + attention_mask
-            attn_weights = torch.max(
-                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
-            )
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
 
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(
-            attn_weights, dim=-1, dtype=torch.float32
-        ).to(query_states.dtype)
-        attn_output = torch.matmul(attn_weights, value_states)
+    #
+    # xformers-attn start
+    #
 
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
-            raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}"
-            )
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
 
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        # end x-formers vs. not x-formers if-else block
+    # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
+    # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
+    if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
+        # input and output should be of form (bsz, q_len, num_heads, head_dim)
+        attn_output = xformers.ops.memory_efficient_attention(
+            query_states, key_states, value_states, attn_bias=None
+        )
+    else:
+        # input and output should be of form (bsz, q_len, num_heads, head_dim)
+        attn_output = xformers.ops.memory_efficient_attention(
+            query_states,
+            key_states,
+            value_states,
+            # attn_bias=attention_mask,
+            attn_bias=xformers.ops.LowerTriangularMask(),
+        )
 
+    if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, q_len, self.num_heads, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
 
+    #
+    # xformers-attn end
+    #
+
     if self.pretraining_tp > 1:
         attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
         o_proj_slices = self.o_proj.weight.split(
@@ -182,103 +152,4 @@ def xformers_forward(
     else:
         attn_output = self.o_proj(attn_output)
 
-    return attn_output, attn_weights, past_key_value
-
-
-def sdp_attention_forward(
-    self,
-    hidden_states: torch.Tensor,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
-    output_attentions: bool = False,
-    use_cache: bool = False,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    # pylint: disable=duplicate-code
-    bsz, q_len, _ = hidden_states.size()
-
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    (
-        query_states,
-        key_states,
-    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
-        query_states, key_states, cos, sin, position_ids
-    )
-    # [bsz, nh, t, hd]
-
-    if past_key_value is not None:
-        # reuse k, v, self_attention
-        key_states = torch.cat([past_key_value[0], key_states], dim=2)
-        value_states = torch.cat([past_key_value[1], value_states], dim=2)
-
-    past_key_value = (key_states, value_states) if use_cache else None
-
-    # We only apply sdp attention if we don't need to output the whole attention matrix
-    if not output_attentions:
-        with torch.backends.cuda.sdp_kernel():
-            attn_output = torch.nn.functional.scaled_dot_product_attention(
-                query_states,
-                key_states,
-                value_states,
-                attn_mask=attention_mask,
-                is_causal=False,
-            )
-            attn_weights = None
-    else:
-        attn_weights = torch.matmul(
-            query_states, key_states.transpose(2, 3)
-        ) / math.sqrt(self.head_dim)
-
-        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
-            raise ValueError(
-                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
-                f" {attn_weights.size()}"
-            )
-
-        if attention_mask is not None:
-            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
-                raise ValueError(
-                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
-                )
-            attn_weights = attn_weights + attention_mask
-            attn_weights = torch.max(
-                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
-            )
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(
-            attn_weights, dim=-1, dtype=torch.float32
-        ).to(query_states.dtype)
-        attn_output = torch.matmul(attn_weights, value_states)
-
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
-            raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}"
-            )
-
-    attn_output = attn_output.transpose(1, 2)
-    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
-
-    attn_output = self.o_proj(attn_output)
-
-    return attn_output, attn_weights, past_key_value
+    return attn_output, None, past_key_value

src/axolotl/utils/models.py

@@ -103,7 +103,7 @@ def load_model(
             )
 
             LOG.info("patching with flash attention")
-            replace_llama_attn_with_flash_attn()
+            replace_llama_attn_with_flash_attn(packed=cfg.sample_packing)
     elif cfg.is_llama_derived_model and cfg.xformers_attention:
         from axolotl.monkeypatch.llama_attn_hijack_xformers import (
             hijack_llama_attention,
@@ -112,9 +112,7 @@ def load_model(
         LOG.info("patching with xformers attention")
         hijack_llama_attention()
     elif cfg.is_llama_derived_model and cfg.sdp_attention:
-        from axolotl.monkeypatch.llama_attn_hijack_xformers import (
-            hijack_llama_sdp_attention,
-        )
+        from axolotl.monkeypatch.llama_attn_hijack_sdp import hijack_llama_sdp_attention
 
         LOG.info("patching with sdp attention")
         hijack_llama_sdp_attention()