diff --git a/megatron/core/transformer/attention.py b/megatron/core/transformer/attention.py
index c53664d037d..f89259be442 100644
--- a/megatron/core/transformer/attention.py
+++ b/megatron/core/transformer/attention.py
@@ -81,6 +81,13 @@
     except ImportError as e:
         pass
 
+try:
+    from flash_attn.cute import flash_attn_varlen_func as flash_attn4_varlen_func
+
+    HAVE_FA4 = True
+except ImportError:
+    HAVE_FA4 = False
+
 try:
     from flash_mla import flash_mla_with_kvcache, get_mla_metadata
 
@@ -828,7 +835,21 @@ def flash_decode_and_prefill(
                 softmax_scale = self.softmax_scale
             else:
                 softmax_scale = q.shape[-1] ** -0.5
-            if HAVE_FA3:
+            if HAVE_FA4:
+                output_total, _ = flash_attn4_varlen_func(
+                    q,
+                    k,
+                    v,
+                    cu_seqlens_q=cu_seqlens_q,
+                    max_seqlen_q=max_seqlen_q,
+                    max_seqlen_k=max_seqlen_k,
+                    seqused_k=seqlens_k,
+                    page_table=block_table,
+                    softmax_scale=softmax_scale,
+                    causal=True,
+                    num_splits=1,
+                )
+            elif HAVE_FA3:
                 # TODO(ksanthanam): Replace with call to flash_attn_varlen_func once
                 # it accepts block_table
                 output_total = self._flash_attention_3_forward_wrapper(
@@ -898,22 +919,47 @@ def flash_decode_and_prefill(
                     causal=True,
                 )
             else:
-                flash_attn_args = {
-                    "q": q,
-                    "k_cache": k,
-                    "v_cache": v,
-                    "cache_seqlens": seqlens_k,
-                    "causal": True,
-                    "page_table" if HAVE_FA3 else "block_table": block_table,
-                    "num_splits": 0 if not self.batch_invariant_mode else 1,
-                }
-                if HAVE_FA3:
-                    output_total = flash_attn3_with_kvcache(**flash_attn_args)
+                if HAVE_FA4:
+                    if getattr(self, "softmax_scale", None) is not None:
+                        softmax_scale = self.softmax_scale
+                    else:
+                        softmax_scale = q.shape[-1] ** -0.5
+                    # Reshape q from (B, S, H, D) to (B*S, H, D) for varlen interface
+                    q_varlen = q.reshape(-1, q.shape[-2], q.shape[-1])
+                    output_total, _ = flash_attn4_varlen_func(
+                        q_varlen,
+                        k,
+                        v,
+                        cu_seqlens_q=cu_seqlens_q,
+                        max_seqlen_q=tokens_per_request,
+                        max_seqlen_k=max_seqlen_k,
+                        seqused_k=seqlens_k,
+                        page_table=block_table,
+                        softmax_scale=softmax_scale,
+                        causal=True,
+                        num_splits=1,
+                    )
+                    # Reshape back to (B, S, H, D)
+                    output_total = output_total.reshape(
+                        num_requests, tokens_per_request, *output_total.shape[1:]
+                    )
                 else:
-                    assert (
-                        not self.batch_invariant_mode
-                    ), "Batch invariant mode is not supported for flash attention 2"
-                    output_total = flash_attn_with_kvcache(**flash_attn_args)
+                    flash_attn_args = {
+                        "q": q,
+                        "k_cache": k,
+                        "v_cache": v,
+                        "cache_seqlens": seqlens_k,
+                        "causal": True,
+                        "page_table" if HAVE_FA3 else "block_table": block_table,
+                        "num_splits": 0 if not self.batch_invariant_mode else 1,
+                    }
+                    if HAVE_FA3:
+                        output_total = flash_attn3_with_kvcache(**flash_attn_args)
+                    else:
+                        assert (
+                            not self.batch_invariant_mode
+                        ), "Batch invariant mode is not supported for flash attention 2"
+                        output_total = flash_attn_with_kvcache(**flash_attn_args)
 
             # Reshape back to (B*S, 1, H, D) for consistent output shape.
             output_total = output_total.reshape(
@@ -973,8 +1019,8 @@ def forward(
         inference_context = deprecate_inference_params(inference_context, inference_params)
 
         if inference_context and inference_context.is_dynamic_batching():
-            assert HAVE_FA3 or is_fa_min_version(
-                "2.7.3"
+            assert (
+                HAVE_FA4 or HAVE_FA3 or is_fa_min_version("2.7.3")
             ), "flash attn verion v2.7.3 and above is required for dynamic batching."
 
         # hidden_states: [sq, b, h]