Remove default nature of attn_bias

Author: alugorey

aten/src/ATen/native/native_functions.yaml

@@ -14858,7 +14858,7 @@
     MPS: _scaled_dot_product_attention_math_mps
   tags: nondeterministic_seeded
 
-- func: _scaled_dot_product_flash_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
+- func: _scaled_dot_product_flash_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   dispatch:
     CUDA: _scaled_dot_product_flash_attention_cuda
     NestedTensorCUDA: _scaled_dot_product_flash_attention_nestedtensor_cuda
@@ -14874,7 +14874,7 @@
     CompositeExplicitAutograd: _scaled_dot_product_fused_attention_overrideable
   tags: nondeterministic_seeded
 
-- func: _scaled_dot_product_flash_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None) -> (Tensor grad_query, Tensor grad_key, Tensor grad_value)
+- func: _scaled_dot_product_flash_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor? attn_bias, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None) -> (Tensor grad_query, Tensor grad_key, Tensor grad_value)
   device_check: NoCheck
   variants: function
   dispatch:
@@ -14915,13 +14915,13 @@
     CUDA: _scaled_dot_product_cudnn_attention_backward_cuda
   tags: nondeterministic_seeded
 
-- func: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None, Tensor? seqused_k=None, Tensor? alibi_slopes=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
+- func: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None, Tensor? seqused_k=None, Tensor? alibi_slopes=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   variants: function
   dispatch:
     CUDA: _flash_attention_forward
   tags: nondeterministic_seeded
 
-- func: _flash_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None) -> (Tensor, Tensor, Tensor)
+- func: _flash_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor? attn_bias, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None) -> (Tensor, Tensor, Tensor)
   device_check: NoCheck
   variants: function
   dispatch:

aten/src/ATen/native/transformers/cuda/attention.cu

@@ -719,6 +719,7 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, c10::SymInt, c10::SymInt, Tensor, Ten
               v_t,
               std::nullopt,
               std::nullopt,
+              std::nullopt,
               max_seqlen_batch_q,
               max_seqlen_batch_k,
               dropout_p,

aten/src/ATen/native/transformers/cuda/attention_backward.cu

@@ -791,6 +791,7 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> _scaled_dot_product_flash_attenti
     q_t,
     k_t,
     v_t,
+    c10::nullopt,
     out_t,
     logsumexp,
     cumulative_sequence_length_q,

tools/autograd/derivatives.yaml

@@ -2873,15 +2873,15 @@
   output_differentiability: [True, False, False, False]
   query, key, value, attn_bias: _scaled_dot_product_efficient_attention_backward(grad, query, key, value, attn_bias, output, log_sumexp, philox_seed, philox_offset, dropout_p, grad_input_mask, is_causal, scale)
 
-- name: _scaled_dot_product_flash_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
+- name: _scaled_dot_product_flash_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   output_differentiability: [True, False, False, False, False, False, False, False, False]
   query, key, value: _scaled_dot_product_flash_attention_backward_symint(grad, query, key, value, attn_bias, output, logsumexp, cum_seq_q, cum_seq_k, max_q, max_k, dropout_p, is_causal, philox_seed, philox_offset, scale)
 
 - name: _scaled_dot_product_flash_attention_for_cpu(Tensor query, Tensor key, Tensor value, float dropout_p=0.0, bool is_causal=False, *, Tensor? attn_mask=None, float? scale=None) -> (Tensor output, Tensor logsumexp)
   output_differentiability: [True, False]
   query, key, value: _scaled_dot_product_flash_attention_for_cpu_backward(grad, query, key, value, output, logsumexp, dropout_p, is_causal, attn_mask, scale)
 
-- name: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None, Tensor? seqused_k=None, Tensor? alibi_slopes=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
+- name: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None, Tensor? seqused_k=None, Tensor? alibi_slopes=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   output_differentiability: [True, False, False, False, False]
   query, key, value: _flash_attention_backward_symint(grad, query, key, value, attn_bias, output, softmax_logsumexp, cum_seq_q, cum_seq_k, max_q, max_k, dropout_p, is_causal, philox_seed, philox_offset, scale, window_size_left, window_size_right)