Use unfused SDPA for short sequences (q_len <= 128 or kv_len <= 128)

Pull Request resolved: #18651

ATT
ghstack-source-id: 376092495
@exported-using-ghexport

Differential Revision: [D96044308](https://our.internmc.facebook.com/intern/diff/D96044308/)

Author: kimishpatel

.ci/scripts/test_lora.sh

@@ -133,7 +133,8 @@ else
 fi
 
 ### QUANTIZATION & PROGRAM DATA SEPARATION ###
-EXPECTED_QUANT_PREFIX="<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant:
+EXPECTED_QUANT_PREFIX="<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant: me
+<think>
 <think>
 Okay, so I need to calculate 15% of 80."
 EXPECTED_QUANT_LORA_PREFIX="

.ci/scripts/test_lora_multimethod.sh

@@ -67,7 +67,8 @@ EXPECTED_LORA_PREFIX="
 <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant
 To calculate 15% of 80"
 
-EXPECTED_BASE_PREFIX="<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant:
+EXPECTED_BASE_PREFIX="<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant: me
+<think>
 <think>
 Okay, so I need to calculate 15% of 80."
 

extension/llm/custom_ops/op_sdpa.cpp

@@ -412,7 +412,12 @@ Tensor& custom_sdpa_out_impl(
       InvalidArgument,
       output);
 
-  bool use_unfused_sdpa = seq_len == 1;
+  // Quantized GEMM kernels may not handle non-contiguous per-head strides
+  // correctly when seq_dim=ONE and seq_len > 1, so keep the conservative
+  // condition for quantized inputs.
+  bool is_quantized = q.scalar_type() == ScalarType::Char;
+  bool use_unfused_sdpa = (!is_quantized) &&
+      (seq_len <= 128 || num_keys_for_causal_attention <= 128);
   if (use_unfused_sdpa) {
     ET_SWITCH_FLOAT_TYPES(output.scalar_type(), ctx, "sdpa", CTYPE, [&] {
       sdpa::impl::cpu_sdpa<CTYPE>(