[Fix] [AMD] aiter: Fix garbage decode output for non-sliding-window models on the vectorized_5d KV layout (#43335)
Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
This commit is contained in:
co-authored by
Cursor
Xiaoyu Zhang
parent
3c07ff2654
commit
1140a8dfd9
@@ -540,6 +540,14 @@ class AiterAttnBackend(AttentionBackend):
|
||||
"SGLANG_USE_AITER_UNIFIED_ATTN"
|
||||
)
|
||||
|
||||
# Decode metadata format. unified_attention and pa_decode_gluon (the
|
||||
# SHUFFLE 5D decode kernel) both read kv_indices as a 2-D
|
||||
# (bs, max_num_blocks_per_seq) page table; the remaining decode kernels
|
||||
# read flat per-token kv_indices addressed through kv_indptr.
|
||||
self.decode_uses_page_table = (
|
||||
self.use_triton_unified_attention or self.kv_cache_is_vectorized_5d
|
||||
)
|
||||
|
||||
# When topk == 1 the EAGLE draft chain is linear, so target_verify's
|
||||
# mask reduces to pure causal and can go through unified_attention
|
||||
# instead of the legacy triton extend_attention_fwd. Gated on non-MLA
|
||||
@@ -2052,7 +2060,7 @@ class AiterAttnBackend(AttentionBackend):
|
||||
kv_indptr[1 : bs + 1] = torch.cumsum(forward_batch.seq_lens, dim=0)
|
||||
kv_indptr = kv_indptr[: bs + 1]
|
||||
|
||||
if not self.use_triton_unified_attention:
|
||||
if not self.decode_uses_page_table:
|
||||
kv_indices = self._get_kv_indices_scratch(
|
||||
forward_batch.seq_lens_sum, forward_batch.seq_lens.device
|
||||
)
|
||||
@@ -2117,7 +2125,7 @@ class AiterAttnBackend(AttentionBackend):
|
||||
qo_indptr = self.qo_indptr_unified_decode[: bs + 1]
|
||||
|
||||
else:
|
||||
if self.use_triton_unified_attention and not self.use_mla:
|
||||
if self.decode_uses_page_table and not self.use_mla:
|
||||
bs = spec_info.kv_indptr.shape[0] - 1
|
||||
kv_indices, swa_page_table = (
|
||||
self._build_unified_page_table_from_spec(spec_info, bs)
|
||||
@@ -2817,9 +2825,9 @@ class AiterAttnBackend(AttentionBackend):
|
||||
else:
|
||||
self.cuda_graph_kv_indices = kv_indices_buf
|
||||
|
||||
if self.use_triton_unified_attention:
|
||||
# Keep a distinct page-table buffer for unified attention. Sharing
|
||||
# cuda_graph_kv_indices with non-unified token indices makes
|
||||
if self.decode_uses_page_table:
|
||||
# Keep a distinct page-table buffer for page-table decode. Sharing
|
||||
# cuda_graph_kv_indices with flat per-token indices makes
|
||||
# page-table width ambiguous after the token buffer is expanded.
|
||||
max_num_blocks_per_seq = (
|
||||
self.max_context_len + self.page_size - 1
|
||||
@@ -2964,14 +2972,12 @@ class AiterAttnBackend(AttentionBackend):
|
||||
kv_last_page_len = None
|
||||
max_q_len = None
|
||||
|
||||
if spec_info is None or (
|
||||
self.use_triton_unified_attention and not self.use_mla
|
||||
):
|
||||
if spec_info is None or (self.decode_uses_page_table and not self.use_mla):
|
||||
max_num_blocks_per_seq = (
|
||||
self.max_context_len + self.page_size - 1
|
||||
) // self.page_size
|
||||
|
||||
if not self.use_triton_unified_attention:
|
||||
if not self.decode_uses_page_table:
|
||||
kv_indptr = self.kv_indptr
|
||||
kv_indptr[1 : bs + 1] = torch.cumsum(seq_lens, dim=0)
|
||||
kv_indptr = kv_indptr[: bs + 1]
|
||||
|
||||
@@ -0,0 +1,53 @@
|
||||
"""GSM8K on a dense, non-sliding-window model with the SHUFFLE 5D KV layout.
|
||||
|
||||
The 5D layout decodes through ``pa_decode_gluon``, which reads ``kv_indices`` as a
|
||||
2-D page table. Sliding-window models always get that table because they force
|
||||
unified attention; a dense model at ``page_size > 1`` only gets it when the decode
|
||||
metadata follows the KV layout, so this is the configuration that guards it.
|
||||
"""
|
||||
|
||||
import contextlib
|
||||
import unittest
|
||||
|
||||
from sglang.srt.environ import envs
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.test.ci.ci_register import register_amd_ci
|
||||
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_MODEL_NAME_FOR_TEST,
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_amd_ci(est_time=210, suite="stage-b-test-1-gpu-small-amd-mi35x")
|
||||
|
||||
|
||||
class TestAiterVectorized5DDenseDecode(CustomTestCase, GSM8KMixin):
|
||||
gsm8k_accuracy_thres = 0.7
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = DEFAULT_MODEL_NAME_FOR_TEST
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls._env_stack = contextlib.ExitStack()
|
||||
cls._env_stack.enter_context(envs.SGLANG_USE_AITER.override(True))
|
||||
cls._env_stack.enter_context(
|
||||
envs.SGLANG_AITER_KV_CACHE_LAYOUT.override("vectorized_5d")
|
||||
)
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
other_args=["--attention-backend", "aiter", "--page-size", "64"],
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
cls._env_stack.close()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user