diff --git a/csrc/fs_io.cpp b/csrc/fs_io.cpp
index 6b7c8d7fe1e5..60d3117f24db 100644
--- a/csrc/fs_io.cpp
+++ b/csrc/fs_io.cpp
@@ -78,32 +78,37 @@ inline int _store_block(const char* tmp_path, const char* dest_path,
 }
 
 // Core single-block load: dst/size are raw pointer + byte count. Returns 0
-// on success, or the errno of the failing step on failure. On failure,
-// the source file is removed since a partially-read block should not be reused.
+// on success, or the errno of the failing step on failure. Removes the source
+// file ONLY on a provable short read (the read completed but returned fewer
+// bytes than requested): stores are atomic, so a too-short file is genuine
+// corruption. Open failures and read errors (bytes_read < 0) are
+// transient/ambiguous and leave the file untouched; a close failure after a
+// full read is harmless and does not fail the load.
 inline int _load_block(const char* source_path, char* dst, size_t size,
                        bool use_o_direct) {
   const int o_direct_flag = use_o_direct ? kODirectFlag : 0;
   const int fd = open(source_path, O_RDONLY | o_direct_flag, 0);
   if (fd < 0) {
-    const int err = errno;
-    unlink(source_path);
-    return err;
+    return errno;
   }
 
   const ssize_t bytes_read = read(fd, dst, size);
-  if (bytes_read < 0 || static_cast<size_t>(bytes_read) != size) {
-    const int err = bytes_read < 0 ? errno : EIO;
+  if (bytes_read < 0) {
+    // Transient read error: leave the file untouched.
+    const int err = errno;
     close(fd);
-    unlink(source_path);
     return err;
   }
-
-  if (close(fd) != 0) {
-    const int err = errno;
+  if (static_cast<size_t>(bytes_read) < size) {
+    // Provable short read: the block is genuinely corrupt, so remove it.
+    close(fd);
     unlink(source_path);
-    return err;
+    return EIO;
   }
 
+  // A close error after a successful full read is harmless: the data is
+  // already in the destination buffer, so the load succeeds.
+  close(fd);
   return 0;
 }
 
@@ -312,8 +317,21 @@ static PyObject* batch_load_block(PyObject* /*self*/, PyObject* args) {
   if (failed_index >= 0) {
     // PyErr_SetFromErrnoWithFilename() reads the errno to format exception.
     errno = failure_errno;
-    return PyErr_SetFromErrnoWithFilename(PyExc_OSError,
-                                          source_paths[failed_index]);
+    PyErr_SetFromErrnoWithFilename(PyExc_OSError, source_paths[failed_index]);
+    // Attach the number of blocks that loaded before the failure so the tier
+    // can keep them (partial success). failed_index == count of blocks read OK.
+    PyObject *etype, *evalue, *etb;
+    PyErr_Fetch(&etype, &evalue, &etb);
+    PyErr_NormalizeException(&etype, &evalue, &etb);
+    if (evalue != nullptr) {
+      PyObject* num = PyLong_FromSsize_t(failed_index);
+      if (num != nullptr) {
+        PyObject_SetAttrString(evalue, "num_succeeded", num);
+        Py_DECREF(num);
+      }
+    }
+    PyErr_Restore(etype, evalue, etb);
+    return nullptr;
   }
 
   Py_RETURN_NONE;
diff --git a/tests/v1/kv_offload/tiering/test_async_lookup.py b/tests/v1/kv_offload/tiering/test_async_lookup.py
index c97fc4442e76..c3f0b725d2a4 100644
--- a/tests/v1/kv_offload/tiering/test_async_lookup.py
+++ b/tests/v1/kv_offload/tiering/test_async_lookup.py
@@ -5,6 +5,8 @@
 import threading
 from collections.abc import Iterable
 
+import pytest
+
 from vllm.v1.kv_offload.base import OffloadKey, ReqContext, make_offload_key
 from vllm.v1.kv_offload.tiering.async_lookup import AsyncLookupManager
 
@@ -156,3 +158,69 @@ def test_shutdown_unblocks_worker(self):
         mgr = InMemoryLookupManager()
         mgr.shutdown()
         assert not mgr._thread.is_alive()
+
+    def test_mark_miss_flips_cached_verdict_without_reprobing(self):
+        """Failed-load livelock regression (#49176). After a failed load the
+        tier calls mark_miss(), flipping the cached True to False; every
+        subsequent lookup then returns False (MISS) directly, WITHOUT enqueuing
+        a fresh batch_lookup — that is what makes the request unable to loop.
+        The entry is retained (as False) not dropped, so cleanup()'s reverse
+        index stays consistent; an unknown key is a no-op."""
+        mgr = InMemoryLookupManager(existing_keys={_key(1), _key(2)})
+        ctx = _ctx("reqA")
+        assert mgr.lookup(_key(1), ctx) is None
+        mgr.lookup(_key(2), ctx)
+        mgr.flush()
+        mgr._results_ready.wait()
+        mgr._results_ready.clear()
+        assert mgr.lookup(_key(1), ctx) is True
+
+        # An unknown key must not raise or plant an entry.
+        mgr.mark_miss([_key(99)])
+        assert _key(99) not in mgr._lookup_state
+
+        # The block is still "present" per the backing set, so a re-probe would
+        # wrongly return True again — the verdict must be served from cache as
+        # False, with no re-probe enqueued, and stay False across steps.
+        mgr.mark_miss([_key(1)])
+        assert mgr.lookup(_key(1), ctx) is False
+        assert mgr._lookup_batch == []  # no fresh probe enqueued
+        mgr.flush()
+        assert mgr._lookup_queue.empty()  # nothing posted to the worker
+        assert mgr.lookup(_key(1), ctx) is False
+
+        # Entry retained (now False) with reverse index intact, so cleanup()
+        # (which direct-indexes _lookup_state per reverse-index key) tears down
+        # both structures without raising.
+        assert mgr._lookup_state[_key(1)].result is False
+        assert _key(1) in mgr._req_keys["reqA"]
+        mgr.cleanup("reqA")
+        assert _key(1) not in mgr._lookup_state
+        assert _key(2) not in mgr._lookup_state
+        assert "reqA" not in mgr._req_keys
+        mgr.shutdown()
+
+    def test_enqueue_once_invariant_enforced(self):
+        """A key is enqueued for probing exactly once, so drain_results() may
+        receive at most one result per key. Normal operation resolves a key a
+        single time; a second result for an already-decided key trips the assert
+        that guards the invariant (a silent overwrite could flip a corrected
+        miss back to True and reopen the failed-load livelock)."""
+        mgr = InMemoryLookupManager(existing_keys={_key(1)})
+        ctx = _ctx("reqA")
+
+        # (a) Normal operation: the key is enqueued once and resolved once, with
+        # no second result left pending.
+        assert mgr.lookup(_key(1), ctx) is None
+        mgr.flush()
+        mgr._results_ready.wait()
+        mgr._results_ready.clear()
+        assert mgr.lookup(_key(1), ctx) is True  # decided exactly once
+        assert mgr._pending_results.empty()
+
+        # (b) A stray/duplicate result for the now-decided key violates the
+        # enqueue-once invariant and must trip the assert.
+        mgr._pending_results.put([(_key(1), True)])
+        with pytest.raises(AssertionError):
+            mgr.drain_results()
+        mgr.shutdown()
diff --git a/tests/v1/kv_offload/tiering/test_fs_tier.py b/tests/v1/kv_offload/tiering/test_fs_tier.py
index 2959ac1aa03d..7d850ca62ae6 100644
--- a/tests/v1/kv_offload/tiering/test_fs_tier.py
+++ b/tests/v1/kv_offload/tiering/test_fs_tier.py
@@ -245,6 +245,11 @@ def test_store_then_load_roundtrip(fs_tier):
     tier.submit_load(job_l)
     load_results = drain(tier)
     assert all(r.success for r in load_results)
+    # A successful load must NOT touch the file: the delete path fires only on
+    # a provable short read, so a good block stays on disk (guards against an
+    # over-eager delete regressing to upstream's delete-on-any-error).
+    for k in (key(1), key(2)):
+        assert os.path.exists(tier.file_mapper.get_file_name(k))
     # Blocks stay on disk after load
     assert lookup_and_wait(tier, [key(1), key(2)]) == [
         LookupResult.HIT,
@@ -545,6 +550,187 @@ def test_out_of_bounds_block_id_smoke(fs_tier, monkeypatch, use_c_ext):
     assert not load_results[0].success
 
 
+@pytest.mark.parametrize("use_c_ext", [True, False])
+def test_failed_load_corrects_verdict_and_removes_corrupt_file(
+    fs_tier, monkeypatch, use_c_ext
+):
+    """Failed-load livelock regression, covering the whole contract.
+
+    A successful promotion leaves the cached HIT and the on-disk block intact.
+    A promotion that short-reads a truncated (corrupt) block fails, and in
+    get_finished_jobs() the tier removes the corrupt file (stores are atomic,
+    so a too-short file is genuine corruption) and marks the cached verdict
+    False. The SAME request's next lookup is then a MISS served from cache with
+    NO re-probe, so the scheduler cannot re-issue the doomed promotion.
+    """
+    import vllm.v1.kv_offload.tiering.fs.io as io_mod
+
+    if use_c_ext and not io_mod._HAS_FSIO_C:
+        pytest.skip("fs_io_C extension not built")
+    monkeypatch.setattr(io_mod, "_HAS_FSIO_C", use_c_ext)
+
+    tier, _ = fs_tier
+    tier.submit_store(make_job(1, [key(1)], [0]))
+    assert all(r.success for r in drain(tier))
+    path = tier.file_mapper.get_file_name(key(1))
+
+    ctx = ReqContext(req_id="livelock-req")
+    assert lookup_and_wait(tier, [key(1)], ctx=ctx) == [LookupResult.HIT]
+
+    # A successful promotion must NOT touch the verdict or the file.
+    tier.submit_load(make_job(2, [key(1)], [0], is_promotion=True))
+    results = drain(tier)
+    assert len(results) == 1 and results[0].success
+    assert tier.lookup(key(1), ctx) == LookupResult.HIT
+    assert os.path.exists(path)
+
+    # Truncate below block_size so the next promotion short-reads.
+    with open(path, "wb") as f:
+        f.write(b"x" * 10)
+    tier.submit_load(make_job(3, [key(1)], [0], is_promotion=True))
+    results = drain(tier)  # get_finished_jobs() marks the verdict False here
+    assert len(results) == 1 and not results[0].success
+
+    # Corrupt file removed; the SAME request now misses from cache, no re-probe.
+    assert not os.path.exists(path)
+    lm = tier._lookup_manager
+    assert tier.lookup(key(1), ctx) == LookupResult.MISS
+    assert lm._lookup_batch == []
+
+    # A FRESH request re-probes the tier (no cached verdict) and misses too,
+    # since the corrupt file is gone -- the real batch_lookup re-probe path.
+    fresh = ReqContext(req_id="fresh-after-short-read")
+    assert lookup_and_wait(tier, [key(1)], ctx=fresh) == [LookupResult.MISS]
+
+
+@pytest.mark.parametrize("use_c_ext", [True, False])
+def test_batched_partial_load_failure_keeps_loaded_blocks(
+    fs_tier, monkeypatch, use_c_ext
+):
+    """A batched promotion stops at the first bad block and reports how many
+    loaded before it (#50321). Corrupt the LAST block: the earlier blocks load
+    fine, so the job reports successful_keys for them and marks only the failed
+    tail a miss. The earlier keys stay HIT — including for the same request —
+    while the corrupt block stays a MISS (its file was removed)."""
+    import vllm.v1.kv_offload.tiering.fs.io as io_mod
+
+    if use_c_ext and not io_mod._HAS_FSIO_C:
+        pytest.skip("fs_io_C extension not built")
+    monkeypatch.setattr(io_mod, "_HAS_FSIO_C", use_c_ext)
+
+    tier, _ = fs_tier
+    keys = [key(1), key(2), key(3)]  # last one is the "bad" block
+    tier.submit_store(make_job(1, keys, [0, 1, 2]))
+    assert all(r.success for r in drain(tier))
+    bad_path = tier.file_mapper.get_file_name(key(3))
+
+    ctx = ReqContext(req_id="batch-req")
+    assert lookup_and_wait(tier, keys, ctx=ctx) == [LookupResult.HIT] * 3
+
+    # Corrupt only the last block, then load the whole batch as one job.
+    with open(bad_path, "wb") as f:
+        f.write(b"x" * 10)
+    tier.submit_load(make_job(2, keys, [0, 1, 2], is_promotion=True))
+    results = drain(tier)
+    # (a) the job fails but reports the two blocks that loaded before the bad one.
+    assert len(results) == 1 and not results[0].success
+    assert tuple(results[0].successful_keys) == (key(1), key(2))
+
+    # (b) Only the failed tail is a miss; the loaded blocks stay HIT on the same
+    # request, and nothing was re-probed.
+    lm = tier._lookup_manager
+    assert [tier.lookup(k, ctx) for k in keys] == [
+        LookupResult.HIT,
+        LookupResult.HIT,
+        LookupResult.MISS,
+    ]
+    assert lm._lookup_batch == []
+
+    # (c) A fresh request re-probes: the loaded blocks are still on disk (HIT),
+    # only the corrupt block was removed (MISS).
+    tier.on_request_finished(ctx)
+    fresh = ReqContext(req_id="fresh-batch-req")
+    assert lookup_and_wait(tier, keys, ctx=fresh) == [
+        LookupResult.HIT,
+        LookupResult.HIT,
+        LookupResult.MISS,
+    ]
+
+
+@pytest.mark.parametrize("use_c_ext", [True, False])
+def test_batched_load_first_block_fails_marks_whole_batch(
+    fs_tier, monkeypatch, use_c_ext
+):
+    """When the FIRST block fails, nothing loaded before it: the job reports no
+    successful_keys (None) and the whole batch is marked a miss for the
+    request."""
+    import vllm.v1.kv_offload.tiering.fs.io as io_mod
+
+    if use_c_ext and not io_mod._HAS_FSIO_C:
+        pytest.skip("fs_io_C extension not built")
+    monkeypatch.setattr(io_mod, "_HAS_FSIO_C", use_c_ext)
+
+    tier, _ = fs_tier
+    keys = [key(1), key(2), key(3)]  # first one is the "bad" block
+    tier.submit_store(make_job(1, keys, [0, 1, 2]))
+    assert all(r.success for r in drain(tier))
+
+    ctx = ReqContext(req_id="batch-first-fail")
+    assert lookup_and_wait(tier, keys, ctx=ctx) == [LookupResult.HIT] * 3
+
+    with open(tier.file_mapper.get_file_name(key(1)), "wb") as f:
+        f.write(b"x" * 10)
+    tier.submit_load(make_job(2, keys, [0, 1, 2], is_promotion=True))
+    results = drain(tier)
+    assert len(results) == 1 and not results[0].success
+    # Nothing loaded before the failure -> no partial success reported.
+    assert results[0].successful_keys is None
+    # The whole batch is a miss for this request.
+    assert [tier.lookup(k, ctx) for k in keys] == [LookupResult.MISS] * 3
+
+
+@pytest.mark.parametrize("use_c_ext", [True, False])
+def test_transient_load_failure_leaves_file(fs_tier, monkeypatch, use_c_ext):
+    """A transient host error (here ELOOP on open) is NOT a short read: the job
+    fails but the block file must survive untouched, on both the C and Python
+    paths. Deleting on a transient error would turn a passing hiccup into
+    permanent data loss."""
+    import vllm.v1.kv_offload.tiering.fs.io as io_mod
+
+    if use_c_ext and not io_mod._HAS_FSIO_C:
+        pytest.skip("fs_io_C extension not built")
+    monkeypatch.setattr(io_mod, "_HAS_FSIO_C", use_c_ext)
+
+    tier, _ = fs_tier
+    tier.submit_store(make_job(1, [key(1)], [0]))
+    assert all(r.success for r in drain(tier))
+    path = tier.file_mapper.get_file_name(key(1))
+    with open(path, "rb") as f:
+        original = f.read()
+
+    # Make open() fail with ELOOP (fd < 0) without truncating the block. Not
+    # chmod 000: CI runs as root, which bypasses permission bits, so open()
+    # would succeed and the load would not fail at all.
+    saved = path + ".saved"
+    loop = path + ".loop"
+    os.rename(path, saved)
+    os.symlink(loop, path)
+    os.symlink(path, loop)
+
+    tier.submit_load(make_job(2, [key(1)], [0], is_promotion=True))
+    results = drain(tier)
+    assert len(results) == 1 and not results[0].success
+
+    # The path is left alone: a non-short-read error must not unlink.
+    assert os.path.lexists(path)
+
+    os.unlink(path)
+    os.unlink(loop)
+    os.rename(saved, path)
+    with open(path, "rb") as f:
+        assert f.read() == original
+
+
 # ---------------------------------------------------------------------------
 # KV events
 # ---------------------------------------------------------------------------
diff --git a/tests/v1/kv_offload/tiering/test_obj_tier.py b/tests/v1/kv_offload/tiering/test_obj_tier.py
index 661438dce633..0e86038dd6ac 100644
--- a/tests/v1/kv_offload/tiering/test_obj_tier.py
+++ b/tests/v1/kv_offload/tiering/test_obj_tier.py
@@ -334,6 +334,32 @@ def test_failed_transfer_reported(self):
         assert len(results) == 1
         assert not results[0].success
 
+    def test_failed_load_marks_verdict_negative(self):
+        """Regression for the failed-load livelock on the obj tier: a
+        cached HIT must not survive a failed load of the same key. On the
+        failed promotion the tier marks the verdict False from
+        get_finished_jobs() (drained here) on the scheduler thread; otherwise
+        the scheduler would re-issue the same doomed promotion every step for
+        the life of the request. The mark is served from cache with no
+        re-probe, so even though the mock object is still 'present' the SAME
+        request now resolves to MISS."""
+        ctx = ReqContext(req_id="obj-livelock")
+        self.tier.submit_store(make_job(1, [key(1)], [0]))
+        assert all(r.success for r in drain(self.tier))
+        # Cache a positive verdict: the object is present, so lookup is a HIT.
+        assert lookup_and_wait(self.tier, [key(1)], ctx=ctx) == [LookupResult.HIT]
+
+        # The promotion the HIT triggered fails.
+        self.agent.check_xfer_state = lambda h: "ERR"
+        self.tier.submit_load(make_job(2, [key(1)], [0]))
+        results = drain(self.tier)
+        assert len(results) == 1 and not results[0].success
+
+        # After the failed promotion the SAME request's lookup must resolve to
+        # MISS (verdict marked False) instead of serving the stale HIT — even
+        # though the object itself is still present in the mock store.
+        assert lookup_and_wait(self.tier, [key(1)], ctx=ctx) == [LookupResult.MISS]
+
     def test_pending_transfer_not_returned_until_done(self):
         # First poll returns PROC; second poll returns DONE.
         call_count = [0]
diff --git a/tests/v1/kv_offload/tiering/test_tiering_offloading.py b/tests/v1/kv_offload/tiering/test_tiering_offloading.py
index 78be7b5f4228..2e9fa224d241 100644
--- a/tests/v1/kv_offload/tiering/test_tiering_offloading.py
+++ b/tests/v1/kv_offload/tiering/test_tiering_offloading.py
@@ -251,6 +251,55 @@ def _start_request(self, req_context: ReqContext = _CTX):
         if req_context.req_id not in self.manager._req_state:
             self.manager.on_new_request(req_context)
 
+    def test_failed_promotion_finalizes_primary_with_failure(self, manager_setup):
+        """A failed promotion still finalizes the primary slots with
+        success=False; the tier corrects its own verdict."""
+        from unittest.mock import patch
+
+        from vllm.v1.kv_offload.tiering.base import JobResult
+
+        self._start_request()
+        # Register an in-flight promotion job for tier1 by hand.
+        job_id = self.manager._next_job_id()
+        self.manager._transfer_jobs[job_id] = JobMetadata(
+            job_id=job_id,
+            keys=to_keys([1, 2]),
+            block_ids=[0, 1],
+            is_promotion=True,
+            req_context=_CTX,
+        )
+        failed = JobResult(job_id=job_id, success=False)
+        with (
+            patch.object(
+                self.secondary_tier1, "get_finished_jobs", return_value=[failed]
+            ),
+            patch.object(self.primary_tier, "complete_write") as completed,
+        ):
+            self.manager._process_finished_jobs()
+        completed.assert_called_once_with(to_keys([1, 2]), _CTX, False)
+
+    def test_successful_promotion_finalizes_primary_with_success(self, manager_setup):
+        from unittest.mock import patch
+
+        from vllm.v1.kv_offload.tiering.base import JobResult
+
+        self._start_request()
+        job_id = self.manager._next_job_id()
+        self.manager._transfer_jobs[job_id] = JobMetadata(
+            job_id=job_id,
+            keys=to_keys([1]),
+            block_ids=[0],
+            is_promotion=True,
+            req_context=_CTX,
+        )
+        ok = JobResult(job_id=job_id, success=True)
+        with (
+            patch.object(self.secondary_tier1, "get_finished_jobs", return_value=[ok]),
+            patch.object(self.primary_tier, "complete_write") as completed,
+        ):
+            self.manager._process_finished_jobs()
+        completed.assert_called_once_with(to_keys([1]), _CTX, True)
+
     def test_take_events_aggregates_tier_owned_events(self, manager_setup):
         primary_event = OffloadingEvent(to_keys([1]), Medium.CPU, removed=False)
         secondary_event1 = OffloadingEvent(to_keys([2]), Medium.STORAGE, removed=False)
diff --git a/vllm/v1/kv_offload/tiering/async_lookup.py b/vllm/v1/kv_offload/tiering/async_lookup.py
index c75a9604009b..a1bcb44c57e5 100644
--- a/vllm/v1/kv_offload/tiering/async_lookup.py
+++ b/vllm/v1/kv_offload/tiering/async_lookup.py
@@ -34,7 +34,7 @@
 import queue
 import threading
 from abc import ABC, abstractmethod
-from collections.abc import Iterable
+from collections.abc import Collection, Iterable
 from dataclasses import dataclass, field
 
 from vllm.logger import init_logger
@@ -170,8 +170,26 @@ def drain_results(self) -> None:
             for key, result in batch:
                 state = self._lookup_state.get(key)
                 if state is not None:
+                    # A key is enqueued for probing exactly once, so a decided
+                    # verdict must never receive a second result. Enforcing it
+                    # keeps a late/duplicate result from resurrecting a stale
+                    # True and reopening the failed-load livelock.
+                    assert state.result is None, (
+                        "cached key received a second lookup result; the "
+                        "enqueue-once invariant is broken and could reopen the "
+                        "failed-load livelock"
+                    )
                     state.result = result
 
+    def mark_miss(self, keys: Collection[OffloadKey]) -> None:
+        """Force the cached verdict for ``keys`` to False after a failed load, so
+        the scheduler stops re-issuing the doomed promotion (livelock, #49176).
+        Keys with no cached entry are skipped."""
+        for key in keys:
+            state = self._lookup_state.get(key)
+            if state is not None:
+                state.result = False
+
     def cleanup(self, req_id: str) -> None:
         """Remove entries no longer needed by any active request.
 
diff --git a/vllm/v1/kv_offload/tiering/fs/io.py b/vllm/v1/kv_offload/tiering/fs/io.py
index 7c70fbae9f1c..9943c90b5472 100644
--- a/vllm/v1/kv_offload/tiering/fs/io.py
+++ b/vllm/v1/kv_offload/tiering/fs/io.py
@@ -139,9 +139,9 @@ def _load_block(
     block_size: int,
     use_o_direct: bool = True,
 ) -> None:
-    """
-    Load callback: read one KV block from disk. Remove the file on failure.
-    """
+    """Read one KV block from disk; remove the file only on a provable short
+    read (a too-short file is genuine corruption) and leave it untouched on any
+    other error."""
     fd: int | None = None
     view_slice = view.cast("B")[offset : offset + block_size]
     o_direct = O_DIRECT if use_o_direct else 0
@@ -150,15 +150,16 @@ def _load_block(
         fd = os.open(source_path, os.O_RDONLY | o_direct)
         bytes_read = os.readv(fd, [view_slice])
         if bytes_read < block_size:
+            # A failure to remove must not mask the short-read error below.
+            try:
+                os.remove(source_path)
+            except OSError as cleanup_exc:
+                logger.warning(
+                    "Failed to remove short-read file %s: %s",
+                    source_path,
+                    cleanup_exc,
+                )
             raise OSError(f"Short read: expected {block_size} bytes, read {bytes_read}")
-    except Exception:
-        try:
-            os.remove(source_path)
-        except OSError as cleanup_exc:
-            logger.warning(
-                "Failed to remove unreadable file %s: %s", source_path, cleanup_exc
-            )
-        raise
     finally:
         if fd is not None:
             os.close(fd)
@@ -200,7 +201,9 @@ def batch_load_block(
     Load a batch of KV blocks from disk into a shared buffer in one call.
 
     Block i is read from source_paths[i] into view[offsets[i] : offsets[i]+block_size].
-    Raises on first error and removes the offending file.
+    Raises on first error (see _load_block for the delete-on-short-read policy).
+    On failure the raised OSError carries ``num_succeeded`` = the number of
+    blocks loaded before the failing one, so the tier can keep them.
     """
     _validate_offsets(view, offsets, block_size)
 
@@ -209,5 +212,11 @@ def batch_load_block(
         view_slices = [view_B[x : x + block_size] for x in offsets]
         return batch_load_block_C(paths, view_slices, use_o_direct)
     else:
-        for path, offset in zip(paths, offsets):
-            _load_block(path, view, offset, block_size, use_o_direct)
+        for i, (path, offset) in enumerate(zip(paths, offsets)):
+            try:
+                _load_block(path, view, offset, block_size, use_o_direct)
+            except OSError as exc:
+                # Blocks 0..i-1 loaded fine; record the count for partial keep.
+                # The C path sets the same attribute via PyObject_SetAttrString.
+                exc.num_succeeded = i  # type: ignore[attr-defined]
+                raise
diff --git a/vllm/v1/kv_offload/tiering/fs/manager.py b/vllm/v1/kv_offload/tiering/fs/manager.py
index 1c8d8b26676c..560e41093a81 100644
--- a/vllm/v1/kv_offload/tiering/fs/manager.py
+++ b/vllm/v1/kv_offload/tiering/fs/manager.py
@@ -148,6 +148,16 @@ def __init__(
                 )
         # Keys of in-flight store jobs, tracked only when events are enabled.
         self._store_job_keys: dict[JobId, list[OffloadKey]] = {}
+        # Keys of in-flight load (promotion) jobs, so a failed load can mark
+        # its own cached lookup verdicts False (see get_finished_jobs).
+        self._load_job_keys: dict[JobId, list[OffloadKey]] = {}
+        # Per load job: how many blocks loaded before a failure (partial keep).
+        # Written by the pool worker inside the load task before it raises (so
+        # before task_done publishes the job); read on the scheduler thread in
+        # get_finished_jobs only for job ids the finished queue returned. Under
+        # the GIL that read cannot observe the finished job without the prior
+        # write, so no extra lock is needed (get_finished is itself lock-free).
+        self._load_progress: dict[JobId, int] = {}
 
         # Extract block size from primary view
         assert primary_kv_view.strides is not None, (
@@ -219,22 +229,47 @@ def submit_store(self, job_metadata: JobMetadata) -> None:
 
     @override
     def submit_load(self, job_metadata: JobMetadata) -> None:
-        task = functools.partial(
-            batch_load_block,
-            [self.file_mapper.get_file_name(key) for key in job_metadata.keys],
-            self._primary_kv_view,
-            [int(bid) * self._block_size for bid in job_metadata.block_ids],
-            self._block_size,
-            self._use_o_direct,
-        )
+        job_id = job_metadata.job_id
+        # Track this load's keys so a failed promotion can mark only its failed
+        # keys as a miss (see get_finished_jobs).
+        keys = list(job_metadata.keys)
+        self._load_job_keys[job_id] = keys
+        paths = [self.file_mapper.get_file_name(key) for key in keys]
+        offsets = [int(bid) * self._block_size for bid in job_metadata.block_ids]
+
+        def load_task() -> None:
+            try:
+                batch_load_block(
+                    paths,
+                    self._primary_kv_view,
+                    offsets,
+                    self._block_size,
+                    self._use_o_direct,
+                )
+            except OSError as exc:
+                # Runs on the pool worker thread. Record how many blocks loaded
+                # before the failure so get_finished_jobs can keep them; this
+                # write precedes task_done, so the scheduler reads it safely
+                # under the GIL once the finished queue hands back this job.
+                num_succeeded = getattr(exc, "num_succeeded", 0)
+                self._load_progress[job_id] = num_succeeded
+                # Surfaces errno (e.g. EMFILE "Too many open files") for both
+                # the C and Python load paths.
+                logger.debug(
+                    "Load of %d blocks for job %s failed at block %d: %s",
+                    len(paths),
+                    job_id,
+                    num_succeeded,
+                    exc,
+                )
+                raise
 
-        self._pool.enqueue_load(job_metadata.job_id, 1, [task])
+        self._pool.enqueue_load(job_id, 1, [load_task])
 
     @override
     def get_finished_jobs(self) -> Iterable[JobResult]:
-        """
-        Collect completed jobs from the finished-jobs queue.
-        """
+        """Collect finished jobs; a failed promotion marks only its failed keys
+        as a miss here (scheduler thread)."""
         results = []
         for job_id, success in self._pool.get_finished():
             if self.events is not None:
@@ -248,7 +283,25 @@ def get_finished_jobs(self) -> Iterable[JobResult]:
                             locality=self.locality,
                         )
                     )
-            results.append(JobResult(job_id=job_id, success=success))
+            load_keys = self._load_job_keys.pop(job_id, None)
+            num_succeeded = self._load_progress.pop(job_id, 0)
+            if load_keys is not None and not success:
+                # A batched load stops at the first bad block and reports how
+                # many loaded before it. Those earlier blocks are kept in the
+                # primary tier (reported via successful_keys); only this block
+                # and the ones after it are marked a miss and recomputed.
+                successful = load_keys[:num_succeeded]
+                failed = load_keys[num_succeeded:]
+                self._lookup_manager.mark_miss(failed)
+                results.append(
+                    JobResult(
+                        job_id=job_id,
+                        success=False,
+                        successful_keys=tuple(successful) if successful else None,
+                    )
+                )
+            else:
+                results.append(JobResult(job_id=job_id, success=success))
         return results
 
     @override
diff --git a/vllm/v1/kv_offload/tiering/obj/manager.py b/vllm/v1/kv_offload/tiering/obj/manager.py
index 2dfc2d30fa2b..fcfea6909fd3 100644
--- a/vllm/v1/kv_offload/tiering/obj/manager.py
+++ b/vllm/v1/kv_offload/tiering/obj/manager.py
@@ -141,6 +141,9 @@ def __init__(
                 )
         # Keys of in-flight store jobs, tracked only when events are enabled.
         self._store_job_keys: dict[JobId, list[OffloadKey]] = {}
+        # Keys of in-flight load (promotion) jobs, so a failed download can
+        # mark its own cached lookup verdicts False (see get_finished_jobs).
+        self._load_job_keys: dict[JobId, list[OffloadKey]] = {}
 
         agent_config = nixl_agent_config(backends=[])
         self._agent = nixl_agent("ObjAgent", agent_config)
@@ -281,6 +284,7 @@ def submit_store(self, job_metadata: JobMetadata) -> None:
         )
 
     def submit_load(self, job_metadata: JobMetadata) -> None:
+        self._load_job_keys[job_metadata.job_id] = list(job_metadata.keys)
         obj_keys = (self._file_mapper.get_file_name(k) for k in job_metadata.keys)
         self._submit_transfer(
             job_metadata.job_id, job_metadata.block_ids, obj_keys, NIXL_READ
@@ -340,12 +344,13 @@ def _poll_active_transfers(self) -> None:
             self._pending_results.append(JobResult(job_id=job_id, success=success))
 
     def get_finished_jobs(self) -> Iterable[JobResult]:
-        """Poll in-flight transfers; return completed (job_id, success) pairs."""
+        """Poll transfers; a failed promotion marks its cached verdicts False
+        here (scheduler thread)."""
         self._poll_active_transfers()
         results = self._pending_results
         self._pending_results = []
-        if self.events is not None:
-            for result in results:
+        for result in results:
+            if self.events is not None:
                 keys = self._store_job_keys.pop(result.job_id, None)
                 if result.success and keys:
                     self.events.append(
@@ -356,6 +361,17 @@ def get_finished_jobs(self) -> Iterable[JobResult]:
                             locality=self.locality,
                         )
                     )
+            # Mark only the keys that did not load as a miss; the request
+            # recomputes them. The miss is per-request (cleared when the request
+            # finishes), so other requests still HIT the blocks that loaded
+            # fine. nixl reports the batch as a whole (successful_keys is None),
+            # so today this marks all keys; the subtraction keeps it correct if
+            # partial results are ever reported.
+            load_keys = self._load_job_keys.pop(result.job_id, None)
+            if load_keys is not None and not result.success:
+                successful = set(result.successful_keys or ())
+                failed = [k for k in load_keys if k not in successful]
+                self._lookup_manager.mark_miss(failed)
         return results
 
     def take_events(self) -> Iterable[OffloadingEvent]: