[V0 deprecation] Remove _VLLM_V1 suffixes from attention backend names (#25489)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
This commit is contained in:
Matthew Bonanni
2025-09-25 17:37:50 +00:00
committed by GitHub
parent 71b25b0d48
commit 3468f17ebe
42 changed files with 131 additions and 174 deletions
@@ -577,8 +577,8 @@ class NixlConnectorWorker:
use_mla=self.use_mla)
self.backend_name = backend.get_name()
attn_backend = backend_name_to_enum(self.backend_name)
self._use_flashinfer = attn_backend == _Backend.FLASHINFER_VLLM_V1
self._use_pallas_v1 = attn_backend == _Backend.PALLAS_VLLM_V1
self._use_flashinfer = attn_backend == _Backend.FLASHINFER
self._use_pallas = attn_backend == _Backend.PALLAS
self.kv_cache_layout = get_kv_cache_layout()
logger.debug("Detected attention backend %s", self.backend_name)
logger.debug("Detected kv cache layout %s", self.kv_cache_layout)
@@ -749,7 +749,7 @@ class NixlConnectorWorker:
# (roughly 8KB vs 5KB).
# Conversely for FlashInfer, K and V are registered in the same region
# to better exploit the memory layout (ie num_blocks is the first dim).
split_k_and_v = not (self.use_mla or self._use_pallas_v1
split_k_and_v = not (self.use_mla or self._use_pallas
or self._use_flashinfer)
tensor_size_bytes = None
for layer_name, cache_or_caches in xfer_buffers.items():
@@ -938,7 +938,7 @@ class NixlConnectorWorker:
tp_ratio = divide(self._tp_size[self.engine_id],
self._tp_size[engine_id])
assert tp_ratio > 0, "Decode TP cannot be smaller than prefill TP"
assert not self._use_pallas_v1 or tp_ratio == 1, \
assert not self._use_pallas or tp_ratio == 1, \
"TPU (pallas_v1) DOES NOT support heterogeneous TP yet."
# Handle tp_size>num_kv_heads: replicate KV cache.