[HiCache]Support hybrid pool staged H2D kernel (#28434)

Co-authored-by: hzh0425 <hzh0425@apache.org>
This commit is contained in:
huangtingwei
2026-06-19 09:48:03 +08:00
committed by GitHub
co-authored by hzh0425
parent 05ee93c44f
commit 6b7ecca663
12 changed files with 1319 additions and 82 deletions
@@ -30,8 +30,8 @@ def _assert_pp_decode_cached_tokens(result, history_len, output_len, label):
class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase):
"""Qwen3-32B + HiCache + PP + UnifiedRadixCache."""
hicache_io_backend = "direct"
hicache_mem_layout = "page_first_direct"
hicache_io_backend = "kernel"
hicache_mem_layout = "page_first"
max_running_requests = 2
kl_threshold = 0.005
gsm8k_threshold = 0.7
@@ -19,8 +19,8 @@ QWEN3_30B_MODEL = "Qwen/Qwen3-30B-A3B-FP8"
class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase):
"""Qwen3-30B-A3B-FP8 + HiCache + CP + UnifiedRadixCache."""
hicache_io_backend = "direct"
hicache_mem_layout = "page_first_direct"
hicache_io_backend = "kernel"
hicache_mem_layout = "page_first"
max_running_requests = 32
kl_threshold = 0.005
gsm8k_threshold = 0.7
@@ -146,9 +146,9 @@ class TestUnifiedDeepSeekV4FlashHiCacheL3(AccuracyTwoPassMixin, CustomTestCase):
"--hicache-storage-prefetch-policy",
"wait_complete",
"--hicache-io-backend",
"direct",
"kernel",
"--hicache-mem-layout",
"page_first_direct",
"page_first",
"--hicache-storage-backend",
"file",
"--swa-full-tokens-ratio",
@@ -210,9 +210,9 @@ class TestUnifiedDeepSeekV4FlashEagleHiCacheL3(AccuracyTwoPassMixin, CustomTestC
"--hicache-storage-prefetch-policy",
"wait_complete",
"--hicache-io-backend",
"direct",
"kernel",
"--hicache-mem-layout",
"page_first_direct",
"page_first",
"--hicache-storage-backend",
"file",
"--enable-cache-report",
@@ -108,9 +108,9 @@ class TestUnifiedMambaHiCache(UnifiedRadixTreeTestMixin, CustomTestCase):
"--hicache-write-policy",
"write_through",
"--hicache-io-backend",
"direct",
"kernel",
"--hicache-mem-layout",
"page_first_direct",
"page_first",
"--max-total-tokens",
"12000",
"--max-mamba-cache-size",
@@ -169,9 +169,9 @@ class TestUnifiedMambaHiCacheL3(AccuracyTwoPassMixin, CustomTestCase):
"--hicache-storage-prefetch-policy",
"wait_complete",
"--hicache-io-backend",
"direct",
"kernel",
"--hicache-mem-layout",
"page_first_direct",
"page_first",
"--hicache-storage-backend",
"file",
"--max-mamba-cache-size",
@@ -170,9 +170,9 @@ class TestGLM5HiRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
"--hicache-storage-prefetch-policy",
"wait_complete",
"--hicache-io-backend",
"direct",
"kernel",
"--hicache-mem-layout",
"page_first_direct",
"page_first",
"--hicache-storage-backend",
"file",
],
@@ -218,9 +218,9 @@ class TestGLM5UnifiedRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
"--hicache-storage-prefetch-policy",
"wait_complete",
"--hicache-io-backend",
"direct",
"kernel",
"--hicache-mem-layout",
"page_first_direct",
"page_first",
"--hicache-storage-backend",
"file",
],