minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD][DSV4] feat: enable DSpark with fp8 unified_kv on gfx950 (#38901)

Browse Source 
Co-authored-by: HAI <hixiao@gmail.com>
This commit is contained in:
amd-danli103
2026-09-20 01:16:39 -07:00
committed by GitHub
co-authored by HAI
parent a8a4d86be9
commit e54009240a
6 changed files with 333 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py
+27 -8
View File
@@ -78,6 +78,19 @@ def get_swa_ring_size(sliding_window: int, is_speculative: bool = False) -> int:
return sliding_window + spec_extra
def resolve_unified_kv_fp8(unified_fp8: Optional[bool] = None) -> bool:
"""Per-pool fp8 layout. None follows SGLANG_DSV4_UNIFIED_KV_FP8.
A caller may pass False so this pool keeps the bf16 ring while the env
stays on (target fused-Q still keys off the global switch).
"""
from sglang.kernels.ops.attention.dsv4.unified_kv_kernels.env_gate import (
is_unified_kv_fp8,
)
return is_unified_kv_fp8() if unified_fp8 is None else bool(unified_fp8)
def _num_dsv4_physical_kv_pages(
size: int, physical_page_size: int, logical_page_size: int
) -> int:
@@ -867,6 +880,7 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
enable_hisparse: bool = False,
online_mtp_max_draft_tokens: int = 0,
num_req_slots: Optional[int] = None,
unified_fp8: Optional[bool] = None,
kv_source_layers: Sequence[int] = (),
full_size: Optional[int] = None,
is_draft_worker: bool = False,
@@ -893,13 +907,6 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
self.compressed_kv_layout_option = compressed_kv_layout
c4_logical_size = c128_size * 32
logger.info(
"Initialize DeepSeekV4TokenToKVPool with "
f"{max_num_reqs=} {swa_size=} {c4_size=} "
f"{c4_logical_size=} {c128_size=} "
f"{c4_state_pool_size=} {c128_state_pool_size=}"
)
self.max_num_reqs = max_num_reqs
# PD preallocation can exceed max_num_reqs;
# the SWA ring must cover every addressable req_pool_idx.
@@ -915,7 +922,19 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
# Resolve the unified-kv gate before any sizing so the two cannot drift.
self._unified_kv = is_unified_kv_triton()
self._unified_kv_fp8 = is_unified_kv_fp8()
self._unified_kv_fp8 = resolve_unified_kv_fp8(unified_fp8)
logger.info(
"Initialize DeepSeekV4TokenToKVPool with "
f"{max_num_reqs=} {swa_size=} {c4_size=} "
f"{c4_logical_size=} {c128_size=} "
f"{c4_state_pool_size=} {c128_state_pool_size=} "
f"unified={self._unified_kv} unified_fp8={self._unified_kv_fp8}"
)
if is_unified_kv_fp8() and not self._unified_kv_fp8:
logger.info(
"SGLANG_DSV4_UNIFIED_KV_FP8 is on; this pool stays bf16 "
"(unified_fp8=False)"
)
# Uniform 512-dim e4m3 layout for the trtllm attention backend
self.uniform_fp8 = (
not self._unified_kv
python/sglang/srt/mem_cache/kv_cache_configurator.py
+16
View File
@@ -133,6 +133,16 @@ def _get_dsv4_compress_state_dtypes() -> tuple[torch.dtype, torch.dtype]:
_is_npu = is_npu()
def unified_fp8_for_dsv4_pool(*, is_draft_worker: bool, spec_algorithm) -> bool:
"""Per-pool fp8 layout. DSpark draft writers scatter bf16, so that pool
stays a bf16 ring; MTP/EAGLE NextN follows the env."""
from sglang.kernels.ops.attention.dsv4.unified_kv_kernels.env_gate import (
is_unified_kv_fp8,
)
return is_unified_kv_fp8() and not (is_draft_worker and spec_algorithm.is_dspark())
def _should_enable_lazy_compaction() -> bool:
"""Lazy compaction default — ON unless
`SGLANG_DISABLE_LAZY_COMPACTION=1` (escape hatch for A/B / rollback).
@@ -1365,6 +1375,11 @@ class KVCacheConfigurator:
kv_layout=kv_layout, compressed_kv_layout=compressed_kv_layout
)
unified_fp8 = unified_fp8_for_dsv4_pool(
is_draft_worker=self.is_draft_worker,
spec_algorithm=self.spec_algorithm,
)
token_to_kv_pool = pool_cls(
max_num_reqs=max_running_requests,
# SWA ring is indexed by req_pool_idx; PD decode inflates req_to_token
@@ -1392,6 +1407,7 @@ class KVCacheConfigurator:
end_layer=self.layer_info.end_layer,
enable_hisparse=get_memory().enable_hisparse,
online_mtp_max_draft_tokens=(max_speculative_num_draft_tokens() or 0),
unified_fp8=unified_fp8,
kv_source_layers=kv_source_layers,
full_size=full_max_total_num_tokens,
**({"is_draft_worker": self.is_draft_worker} if not _is_npu else {}),
python/sglang/srt/model_executor/pool_configurator.py
+27 -6
View File
@@ -968,6 +968,9 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
is the request-scoped fixed pools that do not scale with full_token.
"""
# object.__new__ stubs (SWA floor tests) skip __init__
_dspark_draft_on_bf16 = False
def __init__(self, kvc: KVCacheConfigurator):
self.kv_cache_dtype_str = kvc.kv_cache_dtype_str
cfg = kvc.model_config
@@ -986,6 +989,11 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
# Resolve the unified-kv gate before any sizing so the two cannot drift.
self._unified = is_unified_kv_triton()
self._unified_fp8 = is_unified_kv_fp8()
# DSpark draft still allocates a bf16 ring; target fp8 * (T+1)/T would
# under-count that ring (640 vs 1024). MTP keeps the old inflation.
self._dspark_draft_on_bf16 = bool(
self._unified_fp8 and kvc.spec_algorithm.is_dspark()
)
# Row width across both unified pools: 1024 B bf16, 640 B fp8.
self._unified_row_bytes = dsv4_unified_row_bytes(
self.qk_nope_head_dim, self.qk_rope_head_dim, self._unified_fp8
@@ -1395,20 +1403,32 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
def _fixed_swa_bytes(self, max_running_requests: int) -> int:
"""Unified_kv SWA is a fixed per-request ring, sized by concurrency
(num_req_slots) rather than by full_token. Return its byte footprint
across all full layers, inflated for the draft worker the same way as the
per-token coeff. Returns 0 on the non-unified path (where SWA is already
accounted per-token)."""
(num_req_slots) rather than by full_token. MTP inflates the target ring
by _spec_infl; DSpark+fp8 adds a bf16 draft ring instead (640 vs 1024).
Returns 0 on the non-unified path (SWA already counted per-token)."""
if not self._unified:
return 0
num_req_slots = self._get_num_req_slots(max_running_requests)
ring_bytes = (
target_ring = (
num_req_slots
* self._swa_ring_size
* self._unified_row_bytes
* self.num_layers_total
)
return int(ring_bytes * self._spec_infl)
if self._dspark_draft_on_bf16:
from sglang.srt.mem_cache.deepseek_v4_memory_pool import (
dsv4_unified_row_bytes,
)
draft_row = dsv4_unified_row_bytes(
self.qk_nope_head_dim, self.qk_rope_head_dim, fp8=False
)
# 1 layer is what the shipped DSpark drafts allocate. A multi-stage
# draft would under-count by ~9 MB/layer (128-wide window, ~65 req
# slots), which the (T+1)/T on bytes_per_full_token already covers.
draft_ring = num_req_slots * self._swa_ring_size * draft_row
return int(target_ring + draft_ring)
return int(target_ring * self._spec_infl)
def _to_config(self, sizes: _DSV4PoolSizes) -> MemoryPoolConfig:
full = sizes.full_max_total_num_tokens
@@ -1486,6 +1506,7 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
logger.info(
f"DSV4 memory calculation: unified={self._unified}, "
f"unified_fp8={self._unified_fp8}, "
f"dspark_draft_bf16={self._dspark_draft_on_bf16}, "
f"bytes_per_full_token={self.bytes_per_full_token:.2f}, "
f"available_bytes={available_bytes / (1 << 30):.2f} GB, "
f"c128_state_fixed={c128_state_fixed_bytes / (1 << 30):.2f} GB, "