diff --git a/csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu b/csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
index 7bc435b8e0da..ae93266a4593 100644
--- a/csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
+++ b/csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
@@ -942,9 +942,10 @@ static void launchFullCacheKernel(
 // ────────────────────────────────────────────────────────────────────────────
 // Torch op wrapper
 // ────────────────────────────────────────────────────────────────────────────
-torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
+void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out(
     torch::stable::Tensor const& q_in,           // [N, num_heads_q, 512] bf16
     torch::stable::Tensor const& kv,             // [N, 512] bf16 (read-only)
+    torch::stable::Tensor& q_out,                // [N, q_head_padded, 512]
     torch::stable::Tensor& k_cache,              // [num_blocks, block_bytes] uint8
     torch::stable::Tensor const& slot_mapping,   // [N] int64
     torch::stable::Tensor const& position_ids,   // [N] int64
@@ -970,8 +971,16 @@ torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
   STD_TORCH_CHECK(kv.dim() == 2 && kv.size(1) == 512, "kv shape [N, 512]");
   STD_TORCH_CHECK(q_in.scalar_type() == kv.scalar_type(),
                   "q_in and kv dtype must match");
+  STD_TORCH_CHECK(q_out.device() == q_in.device() && q_out.is_contiguous(),
+                  "q_out must be contiguous and on the same device as q_in");
+  STD_TORCH_CHECK(q_out.scalar_type() == q_in.scalar_type(),
+                  "q_out dtype must match q_in");
   STD_TORCH_CHECK(q_head_padded >= q_in.size(1),
                   "q_head_padded must be >= q_in.size(1) (num_heads_q)");
+  STD_TORCH_CHECK(q_out.dim() == 3 && q_out.size(0) == q_in.size(0) &&
+                      q_out.size(1) == q_head_padded &&
+                      q_out.size(2) == q_in.size(2),
+                  "q_out shape [N, q_head_padded, 512]");
   STD_TORCH_CHECK(k_cache.scalar_type() == torch::headeronly::ScalarType::Byte,
                   "k_cache must be uint8");
   STD_TORCH_CHECK(cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == 64,
@@ -999,11 +1008,6 @@ torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
       q_in.get_device_index());
   const cudaStream_t stream = get_current_cuda_stream(q_in.get_device_index());
 
-  // Allocate the padded q output.  The kernel writes every element (live
-  // region gets RMSNorm+RoPE; pad region gets zeros), so `empty` is safe.
-  auto q_out = torch::stable::new_empty(
-      q_in, {q_in.size(0), q_head_padded, q_in.size(2)}, q_in.scalar_type());
-
   VLLM_STABLE_DISPATCH_HALF_TYPES(
       q_in.scalar_type(), "fused_deepseek_v4_qnorm_rope_kv_insert", [&] {
         using qkv_scalar_t = scalar_t;
@@ -1020,6 +1024,20 @@ torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
                 num_heads_q_padded, cache_block_size_i, kv_block_stride,
                 stream);
       });
+}
+
+torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
+    torch::stable::Tensor const& q_in, torch::stable::Tensor const& kv,
+    torch::stable::Tensor& k_cache,
+    torch::stable::Tensor const& slot_mapping,
+    torch::stable::Tensor const& position_ids,
+    torch::stable::Tensor const& cos_sin_cache, int64_t q_head_padded,
+    double eps, int64_t cache_block_size) {
+  auto q_out = torch::stable::new_empty(
+      q_in, {q_in.size(0), q_head_padded, q_in.size(2)}, q_in.scalar_type());
+  fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out(
+      q_in, kv, q_out, k_cache, slot_mapping, position_ids, cos_sin_cache,
+      q_head_padded, eps, cache_block_size);
   return q_out;
 }
 
diff --git a/csrc/libtorch_stable/ops.h b/csrc/libtorch_stable/ops.h
index e79f22f1d73f..4d0edfa7d21b 100644
--- a/csrc/libtorch_stable/ops.h
+++ b/csrc/libtorch_stable/ops.h
@@ -269,6 +269,14 @@ torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
     torch::stable::Tensor const& cos_sin_cache, int64_t q_head_padded,
     double eps, int64_t cache_block_size);
 
+void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out(
+    torch::stable::Tensor const& q_in, torch::stable::Tensor const& kv,
+    torch::stable::Tensor& q_out, torch::stable::Tensor& k_cache,
+    torch::stable::Tensor const& slot_mapping,
+    torch::stable::Tensor const& position_ids,
+    torch::stable::Tensor const& cos_sin_cache, int64_t q_head_padded,
+    double eps, int64_t cache_block_size);
+
 void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert(
     torch::stable::Tensor& q, torch::stable::Tensor const& kv,
     torch::stable::Tensor& k_cache, torch::stable::Tensor const& slot_mapping,
diff --git a/csrc/libtorch_stable/torch_bindings.cpp b/csrc/libtorch_stable/torch_bindings.cpp
index fec337a8953e..9c5db6c4d290 100644
--- a/csrc/libtorch_stable/torch_bindings.cpp
+++ b/csrc/libtorch_stable/torch_bindings.cpp
@@ -433,6 +433,11 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
       "Tensor q_in, Tensor kv, Tensor! k_cache, "
       "Tensor slot_mapping, Tensor position_ids, Tensor cos_sin_cache, "
       "int q_head_padded, float eps, int cache_block_size) -> Tensor");
+  ops.def(
+      "fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out("
+      "Tensor q_in, Tensor kv, Tensor! q_out, Tensor! k_cache, "
+      "Tensor slot_mapping, Tensor position_ids, Tensor cos_sin_cache, "
+      "int q_head_padded, float eps, int cache_block_size) -> ()");
 
   // FlashInfer V4 full-cache variants: write Q in place (bf16) or to a separate
   // FP8 tensor, and KV into a contiguous 512-wide token-strided cache.
@@ -751,6 +756,8 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
   ops.impl("fused_qk_norm_rope", TORCH_BOX(&fused_qk_norm_rope));
   ops.impl("fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert",
            TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert));
+  ops.impl("fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out",
+           TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out));
   ops.impl(
       "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert",
       TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert));
diff --git a/tests/kernels/test_compressor_kv_cache.py b/tests/kernels/test_compressor_kv_cache.py
index 57eb493e57f8..d8878a0aea90 100644
--- a/tests/kernels/test_compressor_kv_cache.py
+++ b/tests/kernels/test_compressor_kv_cache.py
@@ -20,6 +20,7 @@
 
 from vllm import _custom_ops as ops
 from vllm.models.deepseek_v4.common.ops import (
+    compute_global_topk_indices_and_lens,
     dequantize_and_gather_k_cache,
     quantize_and_insert_k_cache,
 )
@@ -34,6 +35,23 @@
 from .test_fused_indexer_q_rope_quant import quantize_to_mxfp4
 
 
+def test_compute_global_topk_reuses_output_buffers():
+    device = "cuda"
+    topk_indices = torch.tensor(
+        [[0, 3, -1], [1, 2, -1]], dtype=torch.int32, device=device
+    )
+    token_to_req = torch.tensor([0, 1], dtype=torch.int32, device=device)
+    block_table = torch.tensor([[5, 7], [11, 13]], dtype=torch.int32, device=device)
+    is_valid = torch.tensor([True, False], device=device)
+    args = (topk_indices, token_to_req, block_table, 2, is_valid)
+    expected = compute_global_topk_indices_and_lens(*args)
+    outputs = tuple(torch.empty_like(tensor) for tensor in expected)
+    actual = compute_global_topk_indices_and_lens(*args, output_buffers=outputs)
+    for result, output, reference in zip(actual, outputs, expected):
+        assert result.data_ptr() == output.data_ptr()
+        torch.testing.assert_close(result, reference)
+
+
 def _ue8m0_reference(x: torch.Tensor, block_size: int, fp8_max: float):
     """PyTorch reference for UE8M0 FP8 quantization (per-block, power-of-2 scale).
 
diff --git a/tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py b/tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
index ed163a0472ad..5572ee89f210 100644
--- a/tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
+++ b/tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
@@ -257,8 +257,18 @@ def test_q_path_matches_reference(num_tokens: int, n_heads: int, padded_heads: i
         num_blocks, bs, HEAD_BYTES, dtype=torch.uint8, device=device
     ).view(num_blocks, -1)
     slot_mapping = torch.full((num_tokens,), -1, dtype=torch.int64, device=device)
-    q_out = _call_fused(
-        q, padded_heads, kv, k_cache, slot_mapping, positions, cos_sin_cache, eps, bs
+    q_out = torch.empty(num_tokens, padded_heads, HEAD_DIM, dtype=dtype, device=device)
+    torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out(
+        q,
+        kv,
+        q_out,
+        k_cache,
+        slot_mapping,
+        positions,
+        cos_sin_cache,
+        padded_heads,
+        eps,
+        bs,
     )
 
     torch.testing.assert_close(q_out[:, :n_heads], q_ref, rtol=1e-2, atol=1e-2)
diff --git a/tests/kernels/test_fused_indexer_q_rope_quant.py b/tests/kernels/test_fused_indexer_q_rope_quant.py
index 6114b7efd6e7..f8bf944a98f0 100644
--- a/tests/kernels/test_fused_indexer_q_rope_quant.py
+++ b/tests/kernels/test_fused_indexer_q_rope_quant.py
@@ -150,6 +150,23 @@ def test_fused_indexer_q_rope_quant_matches_unfused(
     q_quant_ref, weights_ref = _reference(
         positions, q, cos_sin_cache, weights, softmax_scale, head_scale, use_fp4
     )
+    output_buffers: tuple[torch.Tensor, ...] | None = None
+    OUTPUT_BUFFER_TEST_NUM_TOKENS = 7
+    if num_tokens == OUTPUT_BUFFER_TEST_NUM_TOKENS and cache_dtype == torch.float32:
+        if use_fp4:
+            q_ref, q_scale_ref = q_quant_ref
+            output_buffers = (
+                torch.empty_like(q_ref),
+                torch.empty_like(q_scale_ref)
+                .view(torch.uint8)
+                .reshape(num_tokens, N_HEAD, -1),
+                torch.empty_like(weights_ref),
+            )
+        else:
+            output_buffers = (
+                torch.empty_like(q_quant_ref),
+                torch.empty_like(weights_ref),
+            )
     # use_cutedsl=False: force the triton path even when cutedsl is installed
     # by patching the dispatcher's has_cutedsl() binding to return False.
     cutedsl_patch = (
@@ -169,8 +186,17 @@ def test_fused_indexer_q_rope_quant_matches_unfused(
             softmax_scale,
             head_scale,
             use_fp4,
+            output_buffers=output_buffers,
         )
 
+    if output_buffers is not None:
+        if use_fp4:
+            assert q_quant_fused[0].data_ptr() == output_buffers[0].data_ptr()
+            assert q_quant_fused[1].data_ptr() == output_buffers[1].data_ptr()
+        else:
+            assert q_quant_fused.data_ptr() == output_buffers[0].data_ptr()
+        assert weights_fused.data_ptr() == output_buffers[-1].data_ptr()
+
     if use_fp4:
         q_quant_ref, q_scale_ref = q_quant_ref
         q_quant_fused, q_scale_fused = q_quant_fused
diff --git a/vllm/models/deepseek_v4/attention.py b/vllm/models/deepseek_v4/attention.py
index 5d89337ba790..5813d187f9ba 100644
--- a/vllm/models/deepseek_v4/attention.py
+++ b/vllm/models/deepseek_v4/attention.py
@@ -29,6 +29,7 @@
 from vllm.models.deepseek_v4.common.ops.fused_indexer_q import MXFP4_BLOCK_SIZE
 
 if TYPE_CHECKING:
+    from vllm.models.deepseek_v4.eager_scratch import DeepseekV4EagerScratchPool
     from vllm.v1.attention.backends.mla.sparse_swa import (
         DeepseekSparseSWAMetadata,
     )
@@ -181,6 +182,7 @@ def __init__(
         prefix: str,
         topk_indices_buffer: torch.Tensor | None = None,
         aux_stream_list: list[torch.cuda.Stream] | None = None,
+        eager_scratch_pool: "DeepseekV4EagerScratchPool | None" = None,
     ) -> None:
         super().__init__()
         config = vllm_config.model_config.hf_config
@@ -269,6 +271,7 @@ def __init__(
         )
         self.indexer_rotary_emb = self.rotary_emb
         self.topk_indices_buffer = topk_indices_buffer
+        self.eager_scratch_pool = eager_scratch_pool
 
         self.indexer = None
         if self.compress_ratio == 4:
@@ -290,6 +293,7 @@ def __init__(
                 compress_ratio=self.compress_ratio,
                 prefix=f"{prefix}.indexer",
                 aux_stream=indexer_aux_stream,
+                eager_scratch_pool=eager_scratch_pool,
             )
 
         # Will be None on ROCm for now.
@@ -340,6 +344,7 @@ def __init__(
                 rotate=True,
                 prefix=f"{prefix}.compressor",
                 k_cache_prefix=self.prefix,
+                eager_scratch_pool=eager_scratch_pool,
             )
 
     def forward(
@@ -568,10 +573,24 @@ def _fused_qnorm_rope_kv_insert(
         if cache_dtype == torch.uint8:
             # fp8_ds_mla UE8M0 paged path. Horizontally fused:
             #   Q side:  per-head RMSNorm (no weight) + GPT-J RoPE, zero-filling
-            #            the padding head slots; the kernel allocates and returns
-            #            the padded q tensor.
+            #            the padding head slots.
             #   KV side: GPT-J RoPE + UE8M0 FP8 quant + paged cache insert.
             swa_kv_cache_2d = swa_kv_cache.view(swa_kv_cache.shape[0], -1)
+            if self.eager_scratch_pool is not None:
+                q_out = self.eager_scratch_pool.q_out(q.shape[0])
+                torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out(
+                    q,
+                    kv,
+                    q_out,
+                    swa_kv_cache_2d,
+                    swa_metadata.slot_mapping,
+                    positions,
+                    cos_sin_cache,
+                    self.padded_heads,
+                    self.eps,
+                    swa_metadata.block_size,
+                )
+                return q_out
             return torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
                 q,
                 kv,
@@ -620,6 +639,13 @@ def _fused_qnorm_rope_kv_insert(
         )
         return q_fp8
 
+    def _global_topk_output_buffers(
+        self, topk_indices: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor] | None:
+        if self.compress_ratio != 4 or self.eager_scratch_pool is None:
+            return None
+        return self.eager_scratch_pool.global_topk_outputs(topk_indices)
+
     def get_attn_backend(self) -> type[AttentionBackend]:
         return self.backend_cls
 
@@ -699,6 +725,7 @@ def __init__(
         compress_ratio: int = 1,
         prefix: str = "",
         aux_stream: torch.cuda.Stream | None = None,
+        eager_scratch_pool: "DeepseekV4EagerScratchPool | None" = None,
     ):
         super().__init__()
         self.vllm_config = vllm_config
@@ -711,6 +738,7 @@ def __init__(
         self.rope_dim = config.qk_rope_head_dim  # 64
         self.q_lora_rank = q_lora_rank  # 1536
         self.compress_ratio = compress_ratio
+        self.eager_scratch_pool = eager_scratch_pool
         self.use_fp4_kv = self.vllm_config.attention_config.use_fp4_indexer_cache
         logger.info_once(
             "Using %s indexer cache for Lightning Indexer.",
@@ -774,6 +802,7 @@ def __init__(
             prefix=f"{prefix}.compressor",
             k_cache_prefix=self.k_cache.prefix,
             use_fp4_cache=self.use_fp4_kv,
+            eager_scratch_pool=eager_scratch_pool,
         )
 
         self.indexer_op = SparseAttnIndexer(
@@ -834,6 +863,9 @@ def wq_b_and_q_quant():
             # ReplicatedLinear returns (output, bias); bias is None.
             q, _ = self.wq_b(qr)
             q = q.view(-1, self.n_head, self.head_dim)
+            outputs = None
+            if self.eager_scratch_pool is not None and self.use_fp4_kv:
+                outputs = self.eager_scratch_pool.indexer_q_outputs(q.shape[0])
             return fused_indexer_q_rope_quant(
                 positions,
                 q,
@@ -842,6 +874,7 @@ def wq_b_and_q_quant():
                 self.softmax_scale,
                 self.n_head**-0.5,
                 use_fp4=self.use_fp4_kv,
+                output_buffers=outputs,
             )
 
         # compressor returns None and writes K to the indexer KV cache; the
diff --git a/vllm/models/deepseek_v4/common/ops/cache_utils.py b/vllm/models/deepseek_v4/common/ops/cache_utils.py
index c73d39499eb1..dc06dc91b22f 100644
--- a/vllm/models/deepseek_v4/common/ops/cache_utils.py
+++ b/vllm/models/deepseek_v4/common/ops/cache_utils.py
@@ -438,6 +438,7 @@ def compute_global_topk_indices_and_lens(
     block_table: torch.Tensor,
     block_size: int,
     is_valid_token: torch.Tensor,
+    output_buffers: tuple[torch.Tensor, torch.Tensor] | None = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     """Map local topk indices to global KV cache slots and count valid entries.
 
@@ -447,8 +448,15 @@ def compute_global_topk_indices_and_lens(
     3. Masking padding tokens to length 0
     """
     num_tokens = topk_indices.shape[0]
-    global_topk_indices = torch.empty_like(topk_indices)
-    topk_lens = torch.empty(num_tokens, dtype=torch.int32, device=topk_indices.device)
+    if output_buffers is None:
+        global_topk_indices = torch.empty_like(topk_indices)
+        topk_lens = torch.empty(
+            num_tokens, dtype=torch.int32, device=topk_indices.device
+        )
+    else:
+        global_topk_indices, topk_lens = output_buffers
+        assert global_topk_indices.shape == topk_indices.shape
+        assert topk_lens.shape == (num_tokens,)
     _compute_global_topk_indices_and_lens_kernel[(num_tokens,)](
         global_topk_indices,
         global_topk_indices.stride(0),
diff --git a/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py b/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py
index 3ec0c2b29d5d..5aa00174079a 100644
--- a/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py
+++ b/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py
@@ -295,6 +295,7 @@ def fused_indexer_q_rope_quant(
     index_weights_softmax_scale: float,
     index_weights_head_scale: float,
     use_fp4: bool = False,
+    output_buffers: tuple[torch.Tensor, ...] | None = None,
 ) -> tuple[
     torch.Tensor | tuple[torch.Tensor, torch.Tensor],
     torch.Tensor,
@@ -332,7 +333,13 @@ def fused_indexer_q_rope_quant(
     num_index_q_heads = index_q.shape[1]
     index_q_head_dim = index_q.shape[2]
 
-    index_weights_out = torch.empty_like(index_weights, dtype=torch.float32)
+    if output_buffers is None:
+        index_weights_out = torch.empty_like(index_weights, dtype=torch.float32)
+    else:
+        expected_num_buffers = 3 if use_fp4 else 2
+        assert len(output_buffers) == expected_num_buffers
+        index_weights_out = output_buffers[-1]
+        assert index_weights_out.shape == index_weights.shape
 
     if use_fp4:
         assert index_q_head_dim % MXFP4_BLOCK_SIZE == 0, (
@@ -340,16 +347,23 @@ def fused_indexer_q_rope_quant(
             f"size {MXFP4_BLOCK_SIZE}"
         )
         num_scale_blocks = index_q_head_dim // MXFP4_BLOCK_SIZE
-        index_q_packed = torch.empty(
-            (num_tokens, num_index_q_heads, index_q_head_dim // 2),
-            dtype=torch.uint8,
-            device=index_q.device,
-        )
-        index_q_scale = torch.empty(
-            (num_tokens, num_index_q_heads, num_scale_blocks),
-            dtype=torch.uint8,
-            device=index_q.device,
-        )
+        packed_shape = (num_tokens, num_index_q_heads, index_q_head_dim // 2)
+        scale_shape = (num_tokens, num_index_q_heads, num_scale_blocks)
+        if output_buffers is None:
+            index_q_packed = torch.empty(
+                packed_shape,
+                dtype=torch.uint8,
+                device=index_q.device,
+            )
+            index_q_scale = torch.empty(
+                scale_shape,
+                dtype=torch.uint8,
+                device=index_q.device,
+            )
+        else:
+            index_q_packed, index_q_scale, _ = output_buffers
+        assert index_q_packed.shape == packed_shape
+        assert index_q_scale.shape == scale_shape
         if has_cutedsl():
             # lazily import, otherwise some tests fail due to CUDA driver init failure.
             from vllm.models.deepseek_v4.nvidia.ops.fused_indexer_q_cutedsl import (
@@ -418,7 +432,11 @@ def fused_indexer_q_rope_quant(
     fp8_dtype = current_platform.fp8_dtype()
     use_fnuz = fp8_dtype == torch.float8_e4m3fnuz
     fp8_max = 224.0 if use_fnuz else 448.0
-    index_q_fp8 = torch.empty_like(index_q, dtype=fp8_dtype)
+    if output_buffers is None:
+        index_q_fp8 = torch.empty_like(index_q, dtype=fp8_dtype)
+    else:
+        index_q_fp8, _ = output_buffers
+        assert index_q_fp8.shape == index_q.shape
     if has_cutedsl():
         # lazily import, otherwise some tests fail due to CUDA driver init failure.
         from vllm.models.deepseek_v4.nvidia.ops.fused_indexer_q_cutedsl import (
diff --git a/vllm/models/deepseek_v4/compressor.py b/vllm/models/deepseek_v4/compressor.py
index 590e748145a4..93ae1e89c919 100644
--- a/vllm/models/deepseek_v4/compressor.py
+++ b/vllm/models/deepseek_v4/compressor.py
@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from dataclasses import dataclass
-from typing import Any, ClassVar, cast
+from typing import TYPE_CHECKING, Any, ClassVar, cast
 
 import torch
 from torch import nn
@@ -35,6 +35,9 @@
     SlidingWindowMLASpec,
 )
 
+if TYPE_CHECKING:
+    from vllm.models.deepseek_v4.eager_scratch import DeepseekV4EagerScratchPool
+
 
 def _prefer_two_stage_compressor() -> bool:
     # Platforms that favor the triton variant of two-stage compressor split.
@@ -226,6 +229,7 @@ def __init__(
         prefix: str = "",
         k_cache_prefix="",
         use_fp4_cache: bool = False,
+        eager_scratch_pool: "DeepseekV4EagerScratchPool | None" = None,
     ):
         super().__init__()
         self.compress_ratio = compress_ratio
@@ -235,6 +239,7 @@ def __init__(
         self.prefix = prefix
         self.k_cache_prefix = k_cache_prefix
         self.use_fp4_cache = use_fp4_cache
+        self.eager_scratch_pool = eager_scratch_pool
 
         config = vllm_config.model_config.hf_config
         self.rope_head_dim = config.qk_rope_head_dim
@@ -428,6 +433,10 @@ def forward(
                 store_full_fp8=store_full_fp8,
                 fp8_scale=fp8_scale,
             )
+            if not self.overlap and self.eager_scratch_pool is not None:
+                extra_kwargs["compress_scratch"] = (
+                    self.eager_scratch_pool.compressor_scratch(num_actual)
+                )
         elif self._use_two_stage_fused_compressor:
             # head=512 cr>=128 (no overlap): two-pass split compressor on the
             # prefill suffix, single-pass on the decode prefix.
diff --git a/vllm/models/deepseek_v4/eager_scratch.py b/vllm/models/deepseek_v4/eager_scratch.py
new file mode 100644
index 000000000000..bc46239bfe34
--- /dev/null
+++ b/vllm/models/deepseek_v4/eager_scratch.py
@@ -0,0 +1,137 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from math import prod
+
+import torch
+
+from vllm.models.deepseek_v4.common.ops.fused_indexer_q import MXFP4_BLOCK_SIZE
+from vllm.utils.math_utils import round_up
+
+
+class DeepseekV4EagerScratchPool:
+    """Model-wide outputs and scratch used inside the attention eager break."""
+
+    _ALIGNMENT = 256
+
+    def __init__(
+        self,
+        max_num_tokens: int,
+        padded_q_heads: int,
+        q_head_dim: int,
+        index_q_heads: int,
+        index_q_head_dim: int,
+        index_topk: int,
+        device: torch.device | str,
+    ) -> None:
+        self.max_num_tokens = max_num_tokens
+        self.index_topk = index_topk
+        self._q = torch.empty(
+            (max_num_tokens, padded_q_heads, q_head_dim),
+            dtype=torch.bfloat16,
+            device=device,
+        )
+
+        fp4_specs = (
+            ((max_num_tokens, index_q_heads, index_q_head_dim // 2), torch.uint8),
+            (
+                (
+                    max_num_tokens,
+                    index_q_heads,
+                    index_q_head_dim // MXFP4_BLOCK_SIZE,
+                ),
+                torch.uint8,
+            ),
+            ((max_num_tokens, index_q_heads), torch.float32),
+        )
+        global_specs = (
+            ((max_num_tokens, index_topk), torch.int32),
+            ((max_num_tokens,), torch.int32),
+        )
+        compressor_specs = (((max_num_tokens, q_head_dim), torch.float32),)
+        # FP4 indexer is C4 only, global mapping after FP4 indexer
+        # compressor scratch is C128 only
+        # so here we use max instead of sum
+        aux_bytes = max(
+            self._packed_size(specs)
+            for specs in (fp4_specs, global_specs, compressor_specs)
+        )
+        storage = torch.empty(aux_bytes, dtype=torch.uint8, device=device)
+
+        self._q_outputs: dict[int, torch.Tensor] = {}
+        fp4_values, fp4_scales, fp4_weights = self._views(storage, fp4_specs)
+        self._fp4_template = (fp4_values, fp4_scales, fp4_weights)
+        self._fp4_outputs: dict[
+            int, tuple[torch.Tensor, torch.Tensor, torch.Tensor]
+        ] = {}
+        global_indices, global_lens = self._views(storage, global_specs)
+        self._global_template = (global_indices, global_lens)
+        self._global_outputs: dict[int, tuple[torch.Tensor, torch.Tensor]] = {}
+        self._compressor_template = self._views(storage, compressor_specs)[0]
+        self._compressor_outputs: dict[int, torch.Tensor] = {}
+        self._storage = storage
+
+    @classmethod
+    def _packed_size(
+        cls, specs: tuple[tuple[tuple[int, ...], torch.dtype], ...]
+    ) -> int:
+        offset = 0
+        for shape, dtype in specs:
+            offset = round_up(offset, cls._ALIGNMENT) + prod(shape) * dtype.itemsize
+        return round_up(offset, cls._ALIGNMENT)
+
+    @classmethod
+    def _views(
+        cls,
+        storage: torch.Tensor,
+        specs: tuple[tuple[tuple[int, ...], torch.dtype], ...],
+    ) -> list[torch.Tensor]:
+        offset = 0
+        views = []
+        for shape, dtype in specs:
+            offset = round_up(offset, cls._ALIGNMENT)
+            num_bytes = prod(shape) * dtype.itemsize
+            views.append(storage[offset : offset + num_bytes].view(dtype).view(shape))
+            offset += num_bytes
+        return views
+
+    def q_out(self, num_tokens: int) -> torch.Tensor:
+        output = self._q_outputs.get(num_tokens)
+        if output is None:
+            output = self._q[:num_tokens]
+            self._q_outputs[num_tokens] = output
+        return output
+
+    def compressor_scratch(self, num_tokens: int) -> torch.Tensor:
+        output = self._compressor_outputs.get(num_tokens)
+        if output is None:
+            output = self._compressor_template[:num_tokens]
+            self._compressor_outputs[num_tokens] = output
+        return output
+
+    def indexer_q_outputs(
+        self,
+        num_tokens: int,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        output = self._fp4_outputs.get(num_tokens)
+        if output is None:
+            values, scales, weights = self._fp4_template
+            output = (
+                values[:num_tokens],
+                scales[:num_tokens],
+                weights[:num_tokens],
+            )
+            self._fp4_outputs[num_tokens] = output
+        return output
+
+    def global_topk_outputs(
+        self, topk_indices: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        num_tokens, topk = topk_indices.shape
+        assert topk == self.index_topk
+        output = self._global_outputs.get(num_tokens)
+        if output is None:
+            indices, lens = self._global_template
+            output = (indices[:num_tokens], lens[:num_tokens])
+            self._global_outputs[num_tokens] = output
+        return output
diff --git a/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py b/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py
index 1848c1930db0..48f12647c395 100644
--- a/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py
+++ b/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py
@@ -748,6 +748,9 @@ def _forward_decode(
                         attn_metadata.block_table[:num_decodes],
                         block_size,
                         is_valid,
+                        output_buffers=self._global_topk_output_buffers(
+                            self.topk_indices_buffer[:num_decode_tokens]
+                        ),
                     )
                 )
                 extra_sparse_indices = global_indices.view(num_decode_tokens, 1, -1)
@@ -837,6 +840,7 @@ def _forward_prefill(
                     attn_metadata.block_table,
                     block_size,
                     swa_metadata.is_valid_token[prefill_token_slice],
+                    output_buffers=self._global_topk_output_buffers(local_topk_indices),
                 )
             )
 
diff --git a/vllm/models/deepseek_v4/nvidia/flashmla.py b/vllm/models/deepseek_v4/nvidia/flashmla.py
index 3169ccfce4b3..29ec2d6f2abf 100644
--- a/vllm/models/deepseek_v4/nvidia/flashmla.py
+++ b/vllm/models/deepseek_v4/nvidia/flashmla.py
@@ -176,6 +176,9 @@ def _forward_decode(
                     attn_metadata.block_table[:num_decodes],
                     block_size,
                     is_valid,
+                    output_buffers=self._global_topk_output_buffers(
+                        self.topk_indices_buffer[:num_decode_tokens]
+                    ),
                 )
                 topk_indices = global_indices.view(num_decode_tokens, 1, -1)
             else:
diff --git a/vllm/models/deepseek_v4/nvidia/model.py b/vllm/models/deepseek_v4/nvidia/model.py
index bbda0093726d..ecd838908da1 100644
--- a/vllm/models/deepseek_v4/nvidia/model.py
+++ b/vllm/models/deepseek_v4/nvidia/model.py
@@ -66,6 +66,7 @@
 )
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.models.deepseek_v4.attention import DeepseekV4Attention
+from vllm.models.deepseek_v4.eager_scratch import DeepseekV4EagerScratchPool
 from vllm.models.deepseek_v4.nvidia.flashinfer_sparse import (
     DeepseekV4FlashInferMLAAttention,
     DeepseekV4FlashInferSM120Attention,
@@ -798,6 +799,7 @@ def __init__(
         prefix,
         topk_indices_buffer: torch.Tensor | None = None,
         aux_stream_list: list[torch.cuda.Stream] | None = None,
+        eager_scratch_pool: DeepseekV4EagerScratchPool | None = None,
     ):
         super().__init__()
 
@@ -810,6 +812,7 @@ def __init__(
             prefix=f"{prefix}.attn",
             topk_indices_buffer=topk_indices_buffer,
             aux_stream_list=aux_stream_list,
+            eager_scratch_pool=eager_scratch_pool,
         )
         self.ffn = DeepseekV4MoE(vllm_config, prefix=f"{prefix}.ffn")
 
@@ -986,6 +989,22 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         # (compressor kv_score, indexer.weights_proj, indexer.compressor
         # kv_score). fused_wqa_wkv stays on the default stream.
         aux_stream_list = [torch.cuda.Stream() for _ in range(3)]
+        padded_heads = _select_dsv4_attn_cls(vllm_config).get_padded_num_q_heads(
+            config.num_attention_heads // get_tensor_model_parallel_world_size()
+        )
+        self.eager_scratch_pool: DeepseekV4EagerScratchPool | None = None
+        if not vllm_config.parallel_config.use_ubatching:
+            # TODO: support dbo if needed
+            # this requires the buffer to have ubatch dim
+            self.eager_scratch_pool = DeepseekV4EagerScratchPool(
+                vllm_config.scheduler_config.max_num_batched_tokens,
+                padded_heads,
+                config.head_dim,
+                config.index_n_heads,
+                config.index_head_dim,
+                config.index_topk,
+                current_platform.device_type,
+            )
 
         # Reserved topk indices buffer for all Indexer layers to reuse.
         self.topk_indices_buffer = torch.empty(
@@ -1011,6 +1030,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 prefix=prefix,
                 topk_indices_buffer=self.topk_indices_buffer,
                 aux_stream_list=aux_stream_list,
+                eager_scratch_pool=self.eager_scratch_pool,
             ),
             prefix=f"{prefix}.layers",
         )
diff --git a/vllm/models/deepseek_v4/nvidia/ops/sparse_attn_compress_cutedsl.py b/vllm/models/deepseek_v4/nvidia/ops/sparse_attn_compress_cutedsl.py
index 4ff4b232d10f..960c516c4522 100644
--- a/vllm/models/deepseek_v4/nvidia/ops/sparse_attn_compress_cutedsl.py
+++ b/vllm/models/deepseek_v4/nvidia/ops/sparse_attn_compress_cutedsl.py
@@ -2097,6 +2097,7 @@ def compress_norm_rope_store_cutedsl(
     store_full_kv: bool = False,
     store_full_fp8: bool = False,
     fp8_scale: torch.Tensor | None = None,
+    compress_scratch: torch.Tensor | None = None,
 ) -> None:
     if compress_ratio == 4:
         # For C4A, the single fused kernel is faster than the two-kernel version.
@@ -2129,11 +2130,15 @@ def compress_norm_rope_store_cutedsl(
         )
     else:
         # For C128, the two-kernel version is faster than the single fused kernel.
-        compressed_kv = torch.empty(
-            (num_actual, head_dim),
-            dtype=torch.float32,
-            device=state_cache.device,
-        )
+        if compress_scratch is None:
+            compressed_kv = torch.empty(
+                (num_actual, head_dim),
+                dtype=torch.float32,
+                device=state_cache.device,
+            )
+        else:
+            assert compress_scratch.shape == (num_actual, head_dim)
+            compressed_kv = compress_scratch
         split_kv_compress_norm_rope_insert_sparse_attn_cutedsl(
             state_cache,
             token_to_req_indices,