Fix DynamicCache.reset() leaving the cache in a corrupt, non-empty state

[tritsystem]

Fix DynamicCache.reset() leaving the cache in a corrupt, non-empty state

What does this PR do?

Cache.reset() is a public method. On a DynamicCache (the default cache for
generative models) it does not empty the cache — it leaves it in a corrupt
state where get_seq_length() still reports the old length and the next
update() concatenates onto stale zeros.

Reproduction (on main)

import torch
from transformers import DynamicCache

c = DynamicCache()
c.update(torch.randn(1, 4, 6, 8), torch.randn(1, 4, 6, 8), layer_idx=0)
print(c.get_seq_length(), c.is_initialized)   # 6 True

c.reset()
print(c.get_seq_length(), c.is_initialized)   # 6 True    <-- expected 0, False

k, _ = c.update(torch.randn(1, 4, 3, 8), torch.randn(1, 4, 3, 8), layer_idx=0)
print(k.shape[-2])                             # 9         <-- expected 3
# k[..., :6, :] is all-zero stale state

Root cause

CacheLayerMixin.reset() zeroes keys / values in place (shape unchanged)
and leaves is_initialized = True:

def reset(self) -> None:
    if self.is_initialized:
        self.keys.zero_()
        self.values.zero_()
    if hasattr(self, "cumulative_length"):
        ...

That is correct for StaticLayer — a pre-allocated buffer whose logical length
is a separate cumulative_length counter. DynamicLayer has no length
counter: its length is keys.shape[-2], and it grows via torch.cat. It does
not override reset(), so after reset() the tensors are zeroed but keep their
old shape and is_initialized stays True → get_seq_length() returns the
stale length and update() does torch.cat([stale_zeros, new_states], dim=-2).

Same for DynamicSlidingWindowLayer (cumulative_length gets reset by the base,
but the K/V tensors and is_initialized do not) and DynamicIndexedLayer (its
reset() calls super().reset() and then zeroes the also-concatenation-grown
indexer_keys in place).

generate() itself is unaffected — it only ever resets a StaticCache — but any
user code that reuses a cache across calls hits this.

Fix

• Add DynamicLayer.reset() that returns the layer to its uninitialised
  state (keys = values = None, is_initialized = False, and
  cumulative_length = 0 when present, covering DynamicSlidingWindowLayer).
• DynamicIndexedLayer.reset(): empty indexer_keys (None +
  is_indexer_initialized = False) instead of zeroing it in place.
• LinearAttentionAndFullAttentionLayer.reset() /
  LinearAttentionAndSlidingWindowAttentionLayer.reset() already delegate to
  DynamicLayer.reset / DynamicSlidingWindowLayer.reset, so they are fixed
  transitively.

After the fix, DynamicCache.reset() gives get_seq_length() == 0,
is_initialized is False, and a subsequent update(3) returns a length-3
sequence with no stale positions. reset() on a fresh cache stays a no-op.

Not in scope (noted)

QuantizedLayer (KIVI-style quantized KV cache) subclasses DynamicLayer and
gets the keys / values / is_initialized reset from this change, but keeps
its _quantized_keys / _quantized_values residual buffers. If you'd like, I can
add a small QuantizedLayer.reset() in this PR or a follow-up.

Tests

tests/utils/test_cache_utils.py::SyntheticCacheTest::test_dynamic_cache_reset_returns_to_empty_state
— fails on main (AssertionError: 6 != 0), passes with this change.

pytest tests/utils/test_cache_utils.py -q -k "SyntheticCacheTest or CacheTest or Cropping"
# 17 passed

ruff check / ruff format --check clean on both files.

Before submitting

• Read the contributor guideline.
• This is a bug fix with a repro and a regression test.
• Did you write any new necessary tests? — yes, one.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 33769974777
Result: success | Grafana metrics are not available yet.