diff --git a/docs/features/preload.md b/docs/features/preload.md
index 5787fdb51c1c..f8a51bb08fbe 100644
--- a/docs/features/preload.md
+++ b/docs/features/preload.md
@@ -128,8 +128,10 @@ state rides the data plane.
    meta device (no storage) and requests the tensors from the daemon on its
    GPU.
 3. Before serving anything, the engine and daemon compare a fingerprint of the
-   cached weights: checkpoint content (hashed from safetensors metadata, so
-   identical weights in different directories still match), model
+   cached weights: checkpoint content (hashed from safetensors metadata plus
+   the first 1 KiB of each tensor, so identical weights in different
+   directories still match and a change in any sampled region also changes
+   the key), model
    architecture, TP/DP size/rank, dtype, quantization method and config, text
    RoPE parameters, model revision, KV cache dtype and quantization skip layers,
    and vLLM version. On any mismatch the engine falls back to disk loading
diff --git a/tests/model_executor/model_loader/test_weight_cache.py b/tests/model_executor/model_loader/test_weight_cache.py
index e9ff7dc17c4e..6b6a2070f956 100644
--- a/tests/model_executor/model_loader/test_weight_cache.py
+++ b/tests/model_executor/model_loader/test_weight_cache.py
@@ -775,6 +775,81 @@ def test_ipc_rope_matching(
             server_task.result(timeout=5)
 
 
+def test_hash_checkpoint_distinguishes_weight_values(tmp_path):
+    # Same layout, different values: a base checkpoint and its fine-tune must
+    # not share a key, or an ipc_cache engine silently maps the daemon's stale
+    # weights (#59647).
+    import torch
+    from safetensors.torch import save_file
+
+    from vllm.model_executor.model_loader.weight_cache.protocol import (
+        hash_checkpoint,
+    )
+
+    for name, value in (("base", 0.0), ("finetuned", 1.0)):
+        (tmp_path / name).mkdir()
+        save_file(
+            {"w": torch.full((4, 4), value)},
+            str(tmp_path / name / "model.safetensors"),
+        )
+    assert hash_checkpoint(str(tmp_path / "base")) != hash_checkpoint(
+        str(tmp_path / "finetuned")
+    )
+
+
+def test_hash_checkpoint_stable_across_directory_copies(tmp_path):
+    # The #54921 property: byte-identical checkpoints under different paths
+    # share a key, so a daemon preloaded from one directory can serve an
+    # engine started on the copy.
+    import torch
+    from safetensors.torch import save_file
+
+    from vllm.model_executor.model_loader.weight_cache.protocol import (
+        hash_checkpoint,
+    )
+
+    (tmp_path / "a").mkdir()
+    save_file(
+        {"w": torch.arange(4096, dtype=torch.float32).reshape(64, 64)},
+        str(tmp_path / "a" / "model.safetensors"),
+    )
+    shutil.copytree(tmp_path / "a", tmp_path / "b")
+    assert hash_checkpoint(str(tmp_path / "a")) == hash_checkpoint(str(tmp_path / "b"))
+
+
+def test_hash_checkpoint_catches_single_tensor_change_among_many(tmp_path):
+    # Every tensor's head contributes to the key, so changing one tensor out
+    # of 64 flips it. The sampled reads run through a thread pool, and 64
+    # tensors exceed the worker count, so this also pins down that the digest
+    # does not depend on read scheduling.
+    import torch
+    from safetensors.torch import save_file
+
+    from vllm.model_executor.model_loader.weight_cache.protocol import (
+        hash_checkpoint,
+    )
+
+    tensors = {f"w{i:03d}": torch.full((8, 8), float(i)) for i in range(64)}
+    (tmp_path / "base").mkdir()
+    save_file(tensors, str(tmp_path / "base" / "model.safetensors"))
+    changed = dict(tensors)
+    changed["w037"] = torch.full((8, 8), -1.0)
+    (tmp_path / "changed").mkdir()
+    save_file(changed, str(tmp_path / "changed" / "model.safetensors"))
+    assert hash_checkpoint(str(tmp_path / "base")) != hash_checkpoint(
+        str(tmp_path / "changed")
+    )
+
+
+def test_hash_checkpoint_without_safetensors_returns_none(tmp_path):
+    from vllm.model_executor.model_loader.weight_cache.protocol import (
+        hash_checkpoint,
+    )
+
+    assert hash_checkpoint(str(tmp_path)) is None
+    assert hash_checkpoint(str(tmp_path / "not-a-dir")) is None
+
+
 def test_ipc_loader_copy_mode_reports_no_external_weight_memory():
     """Copy mode clones the weights into the engine, so nothing is external
     (vllm_config is never touched)."""
diff --git a/vllm/model_executor/model_loader/weight_cache/protocol.py b/vllm/model_executor/model_loader/weight_cache/protocol.py
index 334c13e7b0d6..5161686f5fff 100644
--- a/vllm/model_executor/model_loader/weight_cache/protocol.py
+++ b/vllm/model_executor/model_loader/weight_cache/protocol.py
@@ -20,6 +20,7 @@
 import struct
 import tempfile
 import time
+from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass, fields
 from typing import Any, NamedTuple
 
@@ -280,13 +281,67 @@ def _safetensors_header(path: str) -> bytes:
         return size_bytes + f.read(header_len)
 
 
-def hash_checkpoint(model: str) -> str | None:
-    """Fingerprint checkpoint content from local safetensors metadata.
+_TENSOR_HEAD_BYTES = 1024
+# Bounded concurrency for the sampled reads. On NFS each pread costs one RPC
+# RTT, so sequential sampling of a K3-scale checkpoint (~70k tensors) would
+# take over a minute at 1 ms RTT; 32 in-flight reads keep that at a few
+# seconds without hammering the server.
+_FINGERPRINT_READ_WORKERS = 32
+
+
+def _mix_safetensors_fingerprint(
+    path: str, hasher: Any, pool: ThreadPoolExecutor
+) -> None:
+    """Mix one shard's layout and sampled weight bytes into ``hasher``.
 
-    Hashes each shard's safetensors header so a daemon and an engine pointing
-    at identical weights in different directories produce the same key. Returns
-    None when local safetensors files can't be located (e.g. an undownloaded
-    Hugging Face repo id), leaving the caller to fall back to the model path.
+    The header carries tensor names, dtypes, shapes and offsets but no weight
+    values, so hashing it alone gives a base model and its fine-tune the same
+    key (#59647). Sampling the head of every tensor's data region catches
+    value-only changes within that region at the cost of one small read per
+    tensor, and stays stable for byte-identical copies in different
+    directories. The reads go
+    through ``pool`` but are mixed in sorted-name order, so the digest is
+    identical to sequential sampling.
+    """
+    hasher.update(str(os.path.getsize(path)).encode())
+    with open(path, "rb") as f:
+        size_bytes = f.read(8)
+        (header_len,) = struct.unpack("<Q", size_bytes)
+        header = f.read(header_len)
+        hasher.update(size_bytes)
+        hasher.update(header)
+        try:
+            entries = json.loads(header)
+        except json.JSONDecodeError:
+            return
+        data_start = 8 + header_len
+        plan = []
+        for name in sorted(entries):
+            offsets = entries[name].get("data_offsets")
+            if not offsets:
+                continue
+            plan.append(
+                (
+                    data_start + offsets[0],
+                    min(_TENSOR_HEAD_BYTES, offsets[1] - offsets[0]),
+                )
+            )
+        fd = f.fileno()
+        samples = pool.map(lambda spec: os.pread(fd, spec[1], spec[0]), plan)
+        for sample in samples:
+            hasher.update(sample)
+
+
+def hash_checkpoint(model: str) -> str | None:
+    """Fingerprint checkpoint content from local safetensors files.
+
+    Hashes each shard's metadata plus sampled tensor bytes so a daemon and an
+    engine pointing at identical weights in different directories produce the
+    same key, while a difference within any tensor's sampled head produces a
+    different key.
+    Returns None when local safetensors files can't be located (e.g. an
+    undownloaded Hugging Face repo id), leaving the caller to fall back to the
+    model path.
     """
     if not os.path.isdir(model):
         return None
@@ -298,9 +353,10 @@ def hash_checkpoint(model: str) -> str | None:
     if not files:
         return None
     hasher = safe_hash(b"", usedforsecurity=False)
-    for path in sorted(files, key=os.path.basename):
-        hasher.update(os.path.basename(path).encode())
-        hasher.update(_safetensors_header(path))
+    with ThreadPoolExecutor(max_workers=_FINGERPRINT_READ_WORKERS) as pool:
+        for path in sorted(files, key=os.path.basename):
+            hasher.update(os.path.basename(path).encode())
+            _mix_safetensors_fingerprint(path, hasher, pool)
     return hasher.hexdigest()
 
 
@@ -361,7 +417,7 @@ def from_model_config(
         may mutate hf_config.quantization_config, which would change the hash
         between the daemon and the engine.
 
-        The checkpoint is identified by a hash of its safetensors metadata when
+        The checkpoint is identified by a hash of its safetensors files when
         the weights are available locally, so a daemon and engine referencing
         identical weights in different directories still match; otherwise it
         falls back to the model path.