diff --git a/python/sglang/srt/mem_cache/cp_hicache_trace.py b/python/sglang/srt/mem_cache/cp_hicache_trace.py
index dac1ca7f2..085931115 100644
--- a/python/sglang/srt/mem_cache/cp_hicache_trace.py
+++ b/python/sglang/srt/mem_cache/cp_hicache_trace.py
@@ -196,8 +196,12 @@ def cptrace(level: int, tag: str, **fields) -> None:
 
 
 def _dump_enabled_rid(forward_batch):
-    """Return (dump_dir, rid, cprank) if this forward should dump, else None.
-    Gated by SGLANG_NSA_DUMP_DIR + rid starting 'dump-' + EXTEND forward."""
+    """Return (dump_dir, rid, cprank, mode) if this forward should dump, else None.
+    Two rid-controlled modes (gated by SGLANG_NSA_DUMP_DIR + EXTEND forward):
+      'dumpfh-' -> mode 'fh'  : final_hidden ONLY (tiny), works for FRESH prefills too
+                                 -> the cache-BUST-vs-cache-HIT discriminator.
+      'dump-'   -> mode 'full': lean per-layer trajectory (attn_in/out, q_all), and
+                                 only CACHE-HIT extends (skip huge fresh prefills)."""
     try:
         d = envs.SGLANG_NSA_DUMP_DIR.get()
     except Exception:
@@ -209,20 +213,22 @@ def _dump_enabled_rid(forward_batch):
         return None
     rids = getattr(forward_batch, "rids", None)
     rid = str(rids[0]) if rids else ""
-    if not rid.startswith("dump-"):
+    if rid.startswith("dumpfh-"):
+        mode = "fh"
+    elif rid.startswith("dump-"):
+        mode = "full"
+        # 'full' only on cache-hit extends; fresh full prefills are huge + not the repro.
+        epl = getattr(forward_batch, "extend_prefix_lens_cpu", None)
+        try:
+            if epl is not None and int(epl.sum().item()) <= 0:
+                return None
+        except Exception:
+            pass
+    else:
         return None
-    # Only CACHE-HIT extends (have a reloaded/L1 prefix). Skip fresh full prefills
-    # (extend_prefix_len==0): they dump the whole sequence (huge) and aren't the
-    # repro. This lets `--repeat 2` dump only the L1-hit (rep1) + L2-reload (small).
-    epl = getattr(forward_batch, "extend_prefix_lens_cpu", None)
-    try:
-        if epl is not None and int(epl.sum().item()) <= 0:
-            return None
-    except Exception:
-        pass
     lay = getattr(forward_batch, "cp_shared_kv_layout", None)
     cprank = int(getattr(lay, "cp_rank", -1)) if lay is not None else -1
-    return d, rid, cprank
+    return d, rid, cprank, mode
 
 
 def dump_tensors(forward_batch, layer_id, tag, tensors: dict, *, big_layers=None) -> None:
@@ -231,8 +237,9 @@ def dump_tensors(forward_batch, layer_id, tag, tensors: dict, *, big_layers=None
     (token order, comparable; relerr offline beats fp-nondeterminism). 'big' tensors
     (e.g. kv_cache, passed with big_layers=<spread>) are kept ONLY at those layers to
     bound size; everything else at every layer. Gated via _dump_enabled_rid."""
-    if _dump_enabled_rid(forward_batch) is None:
-        return
+    info = _dump_enabled_rid(forward_batch)
+    if info is None or info[3] != "full":
+        return  # per-layer trajectory only in 'full' mode (cache-hit); 'fh' skips it
     try:
         import torch
 
@@ -259,16 +266,23 @@ def dump_flush(forward_batch, *, final_hidden=None, positions=None) -> None:
     info = _dump_enabled_rid(forward_batch)
     if info is None:
         return
-    d, rid, cprank = info
+    d, rid, cprank, mode = info
     try:
         import os
         import torch
 
         buf = getattr(forward_batch, "_cp_dump_buf", None)
         forward_batch._cp_dump_buf = None
-        payload = {"layers": buf or {}, "rid": rid, "cprank": cprank}
+        payload = {"rid": rid, "cprank": cprank, "mode": mode}
+        if mode == "full":
+            payload["layers"] = buf or {}
         if isinstance(final_hidden, torch.Tensor):
-            payload["final_hidden"] = final_hidden.detach().to("cpu")
+            # 'fh' mode: keep only the LAST token's hidden (the first-token source)
+            # so a fresh full-prefill dump stays tiny; 'full' keeps the extend rows.
+            fh = final_hidden.detach()
+            if mode == "fh" and fh.dim() >= 1 and fh.shape[0] > 1:
+                fh = fh[-1:]
+            payload["final_hidden"] = fh.to("cpu")
         if isinstance(positions, torch.Tensor):
             payload["positions"] = positions.detach().to("cpu")
         for name in ("seq_lens_cpu", "extend_prefix_lens_cpu", "extend_seq_lens_cpu"):
diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 1bdbed169..2ddc4863c 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -2146,17 +2146,6 @@ class DeepseekV2Model(nn.Module):
                 else:
                     hidden_states, _ = self.norm(hidden_states, residual)
 
-        # Complete-dump flush at model end (gated by SGLANG_NSA_DUMP_DIR + rid 'dump-'):
-        # write the per-layer buffer + final hidden + metadata to one file per (rid,rank).
-        try:
-            from sglang.srt.mem_cache.cp_hicache_trace import dump_flush as _cp_dump_flush
-
-            _cp_dump_flush(
-                forward_batch, final_hidden=hidden_states, positions=positions
-            )
-        except Exception:
-            pass
-
         if getattr(forward_batch, "capture_draft_hidden_states", False):
             forward_batch.draft_hidden_states = hidden_states
 
@@ -2172,6 +2161,19 @@ class DeepseekV2Model(nn.Module):
                     forward_batch,
                     torch.cuda.current_stream(),
                 )
+
+        # Dump flush at model end (gated by SGLANG_NSA_DUMP_DIR + rid 'dump-'/'dumpfh-').
+        # Placed AFTER the CP gather so final_hidden is the model's actual output (the
+        # first-token source); for 'fh' mode this is the cache-bust-vs-cache-hit probe.
+        try:
+            from sglang.srt.mem_cache.cp_hicache_trace import dump_flush as _cp_dump_flush
+
+            _cp_dump_flush(
+                forward_batch, final_hidden=hidden_states, positions=positions
+            )
+        except Exception:
+            pass
+
         if len(aux_hidden_states) == 0:
             return hidden_states
         return hidden_states, aux_hidden_states