[Diffusion] Register runtime cache for direct DLO H2D

Avoid recurrent mmap-to-pinned packing in no-AllGather DLO by registering transform-complete runtime-cache mappings under an explicit per-worker budget. Copy mapped tensors directly into the rotating device buffers while preserving bounded-staging fallback and ordered shutdown cleanup.

Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
Hongsheng Liu committed 2026-08-19 11:21:28 +08:00
1 parent 67f15e73cd
commit 23ab14a15b
18 files changed
+976 -74

No files matched your search

@@ -46,7 +46,9 @@ produce a direct-checkpoint mmap plan for a proven-compatible runtime layout;
otherwise it uses the ordinary loader. In no-AllGather mode, an opt-in Phase B
runtime cache can normalize those final ordinary-loader DiT tensors into an
immutable node-local mmap entry. Consequently, replicas can share either
proven-compatible checkpoint pages or equivalent final runtime layouts.
proven-compatible checkpoint pages or equivalent final runtime layouts. An
opt-in CUDA registration budget can make the complete final mapping directly
H2D-copyable, avoiding the recurrent host packing otherwise required by mmap.
The Phase A shared-mmap support boundary is TP1. TP greater than one is an
ordinary-loader compatibility path: DLO can consume the resulting TP-local
@@ -138,12 +140,13 @@ error retains the already-valid ordinary tensors.
DLO maps this plan without moving the DiT to `meta` and without rerunning any
post-load hook. It changes tensor storage while preserving `Parameter` objects,
then runs the existing `gc.collect()`/`malloc_trim()` cleanup before installing
streaming hooks. The source mappings remain open through DLO disable and feed
the existing bounded two-slot host staging path. Because file-backed pages are
not pinned, every streamed block is first packed from mmap storage into one of
those pinned slots and is then copied to the device. This removes model-sized
private pinned storage, but adds a recurrent host-to-host copy and can amplify
TP synchronization waits when several ranks compete for host bandwidth.
streaming hooks. The source mappings remain open through DLO disable. With the
default zero registration budget they feed the existing bounded two-slot host
staging path. Every streamed block is first packed from pageable mmap storage
into one of those pinned slots and is then copied to the device. This removes
model-sized private pinned storage, but adds a recurrent host-to-host copy and
can amplify TP synchronization waits when several ranks compete for host
bandwidth.
This version is deliberately opt-in and steady-state-oriented. It requires
roughly one local-disk model copy per runtime identity, has no eviction, and
@@ -151,6 +154,43 @@ still performs ordinary loading plus full content-hash passes in every rank.
Skip-load-on-hit is a separate follow-up after loader side effects can be
modeled safely.
### Registered runtime-cache transfer
`dlo_runtime_cache_pin_limit_gib > 0` opts a CUDA worker into complete
runtime-cache registration. It also changes loader selection: no-AllGather
uses the ordinary loader and final runtime cache even when a TP1 direct-
checkpoint plan is available. A checkpoint binding can carry a deferred
adapter such as grouped-QKV reordering. Registering that raw source would not
remove the adapter's recurrent CPU allocation/copy, whereas the runtime cache
contains the transform-complete final tensors.
After mapping and validating every final tensor, DLO groups source storages by
backing file, page-aligns their address ranges, and coalesces overlapping or
adjacent ranges within that mapping. The complete byte count is checked
against the per-worker budget before any CUDA call. It then registers every
range with the read-only CUDA host-registration flag and verifies that PyTorch
recognizes every mapped tensor as pinned. Failure rolls back any ranges already
registered and preserves the existing two-slot staging path. Partial direct
transfer is deliberately unsupported. A rollback failure aborts worker
initialization so a still-registered range is retained until process/context
teardown rather than being unsafely unmapped.
On success, the hooks copy each immutable tensor view directly into its offset
in the existing flattened rotating device buffer on the copy stream. Parameter
repointing, two-block HBM residency, and H2D/compute overlap are unchanged. The
implementation currently issues more H2D copy operations than block packing,
but avoids the much larger recurrent CPU copy. A packed cache schema should be
considered only if launch-count measurements justify its added format and
publication complexity.
Registration is per process and CUDA context; physical file-cache pages remain
shared and are not copied into one private host allocation per worker. The
operator must nevertheless budget page-locked memory and satisfy OS/CUDA
registration limits for every worker. During shutdown, DLO first synchronizes
outstanding device work, unregisters all ranges, and then closes the safetensors
mappings. Worker teardown invokes this cleanup before destroying the
distributed environment and CUDA context.
### AllGather path
With the default `dlo_use_allgather=True`, each rank stores approximately
@@ -203,6 +243,10 @@ the persistent private full-model copy per pure-DP process, but each process
still packs and transfers every complete block. Sharing is node-local; each
node has its own page cache.
For a runtime-cache mapping, a positive registration budget replaces those
host slots with direct copies from the shared final tensor views. A zero budget
or failed registration retains the bounded staging behavior.
When direct mmap preflight fails, the regular model loader remains responsible
for preparing each rank's weights, including TP-local tensors or HSDP-managed
parameters. Supported CPU layouts may then enter the opt-in runtime cache.
@@ -225,9 +269,9 @@ This mode means:
| Parallelism | DLO + AllGather | DLO without AllGather |
|---|---|---|
| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages; the optional runtime cache shares equivalent final layouts while excluding DP rank from its identity. |
| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active without a DLO weight collective. Runtime-cache v1 excludes SP rank and includes conservative SP implementation/world-size guards. |
| **TP > 1** | Outside the Phase A direct-mmap scope. The loader preserves TP-local layouts and DLO may apply DP/SP host sharding to ordinary runtime tensors. | The ordinary TP-aware loader produces rank-local tensors. Runtime-cache v1 creates one entry per TP coordinate, shareable by equivalent DP/SP processes. |
| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages; the optional runtime cache shares equivalent final layouts while excluding DP rank from its identity. CUDA workers may register the complete final mapping for direct H2D. |
| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active without a DLO weight collective. Runtime-cache v1 excludes SP rank and includes conservative SP implementation/world-size guards. Registered transfer does not remove ordinary SP activation/attention collectives. |
| **TP > 1** | Outside the Phase A direct-mmap scope. The loader preserves TP-local layouts and DLO may apply DP/SP host sharding to ordinary runtime tensors. | The ordinary TP-aware loader produces rank-local tensors. Runtime-cache v1 creates one entry per TP coordinate, shareable by equivalent DP/SP processes. Registering those final mappings removes staging-driven TP skew but does not remove ordinary TP collectives. |
| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted by configuration. HSDP owns parameter sharding and its own gathers; DLO only stages rank-local parameters. End-to-end coverage is limited. |
### Combined dimensions
@@ -286,6 +330,8 @@ Current source-level validation includes:
rejection;
- loader ordering through final post-load mutation and DLO remapping without
rerunning post-load hooks;
- all-or-nothing page-aligned registration, budget rejection, partial-failure
rollback, direct-copy staging bypass, and explicit shutdown cleanup;
- resident-layer requests requiring no-AllGather;
- DP request-wave validation for denoising-step compatibility;
- sharding, double-buffer, AllGather-size, and heterogeneous-block regression
@@ -397,13 +443,48 @@ rose from 3.75 to 32.49 seconds, while communication-kernel time rose from
summed across ranks and can overlap; they identify the mechanism, not wall
time. Matching private/cache outputs were byte-identical.
The result confirms that runtime-cache v1 is a host-capacity option, not a
latency-neutral replacement for private pinned weights. Broader SP, model,
platform, and production-quality validation remains tracked in
The result confirms that zero-budget runtime-cache staging is a host-capacity
option, not a latency-neutral replacement for private pinned weights. Broader
SP, model, platform, and production-quality validation remains tracked in
[issue #6231](https://github.com/vllm-project/vllm-omni/issues/6231). The
historical B300 rows above predate the runtime cache and must not be treated as
Phase B memory validation.
### L20X registered-runtime-cache smoke
The read-only registered path was then tested with the same MiniMax-H3
workload. These are functional and mechanism-validation measurements, not a
production benchmark. The TP1 pair used one request per wave. Both modes moved
exactly 121.86 GiB H2D in the profiled request and produced byte-identical video
and audio for both measured seeds.
| TP1 no-AllGather mode | Mean wave | Throughput | Request-time host PSS | `Private_Dirty` | Engine peak HBM | CPU `aten::copy_` | GPU compute |
|---|---:|---:|---:|---:|---:|---:|---:|
| Checkpoint mmap + bounded staging | 33.87 s | 0.0295 req/s | 133.65 GiB | 71.16 GiB | 13,226 MB | 21.91 s | 419.28 ms |
| Final runtime cache + registered direct H2D | 4.88 s | 0.2055 req/s | 129.41 GiB | 66.93 GiB | 13,226 MB | 1.64 s | 419.05 ms |
Registration covered 61.73 GiB in 13 page-aligned ranges. Mean wave latency
fell by 85.6% and throughput increased 6.94x. GPU compute and H2D payload were
unchanged, identifying recurrent host transform/packing as the removed work.
The direct path issued 2,425 H2D events versus 1,945 for packed staging, but
that launch increase was much smaller than the eliminated CPU cost.
Two independent TP2 engines also registered the same two TP-coordinate cache
entries across four workers. Compared with the zero-budget cache, mean wave
latency improved from 9.21/12.55 seconds to 3.80/3.86 seconds for engines A/B;
the private-loader reference was 3.61/3.72 seconds. Registered request-time
host PSS was 224.26 GiB versus 232.23 GiB staged and 366.20 GiB private.
Aggregate CPU `aten::copy_` time was 3.83 seconds registered, 32.49 seconds
staged, and 3.75 seconds private; aggregate GPU compute and the 264.24 GiB H2D
payload were effectively unchanged. Ordinary TP communication remained, but
the extra wait caused by staging-driven rank skew largely disappeared.
All workers explicitly unregistered before releasing their mappings, the
system's acquired-minus-released pin counter returned to its pre-run value,
and every GPU allocation was released. Cold runtime-cache publication remained
expensive because it still performs ordinary loading and multiple full-content
passes; that startup problem is outside the registered transfer scope.
## Recommendations
- Use **DP + DLO AllGather** for the supported throughput and host-memory
@@ -413,8 +494,10 @@ Phase B memory validation.
- Use **no-AllGather** when independent replica execution is required. TP1
direct-mmap deployments can share checkpoint pages per node. Enable the
runtime cache only when independent replicas have repeated final layouts and
host capacity is more important than the added mmap-to-pinned staging cost.
Benchmark the target CPU/memory topology, especially for matching
coordinates across multiple TP engines.
benchmark the target CPU/memory topology, especially for matching
coordinates across multiple TP engines. On CUDA, provide an explicit
registration budget large enough for the complete final layout when
recurrent staging cost is unacceptable; leave it at zero when page-locking
that layout is not operationally acceptable.
- Prefer **HSDP alone** for production HSDP deployments until the combined
HSDP + DLO no-AllGather path has broader end-to-end coverage.
@@ -7,7 +7,9 @@ each rank streams a complete block independently. Compatible TP1 deployments
can share checkpoint-backed host pages among processes on the same node;
otherwise DLO streams the ordinary loader's rank-local tensors. The optional
runtime-weight cache extends node-local sharing to final ordinary-loader
layouts, including matching TP coordinates across independent engines.
layouts, including matching TP coordinates across independent engines. On
CUDA, an explicit registration budget can make those shared mappings directly
H2D-copyable and avoid recurrent mmap-to-pinned host packing.
See the [DLO feature design](../../../design/feature/offloader/distributed_layerwise_offload.md)
for the implementation contract and compatibility matrix.
@@ -48,7 +50,8 @@ vllm serve /path/to/model --omni \
--tensor-parallel-size 2 \
--dlo-no-use-allgather \
--dlo-enable-runtime-cache \
--dlo-runtime-cache-dir /var/cache/vllm-omni/dlo-runtime-weights
--dlo-runtime-cache-dir /var/cache/vllm-omni/dlo-runtime-weights \
--dlo-runtime-cache-pin-limit-gib 80
# Sequence parallel deployment
vllm serve /path/to/model --omni \
@@ -65,6 +68,7 @@ omni = Omni(
dlo_use_allgather=False,
dlo_enable_runtime_cache=True,
dlo_runtime_cache_dir="/var/cache/vllm-omni/dlo-runtime-weights",
dlo_runtime_cache_pin_limit_gib=80,
)
```
@@ -79,6 +83,7 @@ omni = Omni(
| `--dlo-enable-runtime-cache` | Normalize final ordinary-loader DiT weights into a node-local mmap cache; requires no-AllGather | `false` |
| `--dlo-runtime-cache-dir PATH` | Shared local-disk cache root | `~/.cache/vllm-omni/dlo-runtime-weights` |
| `--dlo-runtime-cache-lock-timeout SECONDS` | Maximum wait for another process publishing the same layout | `600` |
| `--dlo-runtime-cache-pin-limit-gib GIB` | Per-worker limit for registering the complete read-only runtime-cache mapping for direct H2D; zero keeps bounded host staging | `0` |
| `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |
## Host-weight loading
@@ -95,14 +100,21 @@ DLO may still consume those TP-local tensors, but this is a compatibility path:
it does not share checkpoint-backed runtime weights across DP replicas and
provides no shared-mmap host-memory guarantee.
When `--dlo-enable-runtime-cache` is set with no-AllGather, that fallback gains
a second storage opportunity. Every rank still completes ordinary loading,
loader callbacks, post-load processing, validation, and calibration. It then
hashes the final CPU DiT parameters and persistent buffers. One process per
equivalent runtime layout publishes sharded safetensors plus a checksummed
manifest; peers validate and mmap the same files. DLO replaces only tensor
storage, preserves the final `Parameter` objects, and releases the private
allocator-backed copies before installing its streaming hooks.
When `--dlo-enable-runtime-cache` is set with no-AllGather, ordinary loading
gains a second storage opportunity. Every rank completes loader callbacks,
post-load processing, validation, and calibration. It then hashes the final
CPU DiT parameters and persistent buffers. One process per equivalent runtime
layout publishes sharded safetensors plus a checksummed manifest; peers
validate and mmap the same files. DLO replaces only tensor storage, preserves
the final `Parameter` objects, and releases the private allocator-backed copies
before installing its streaming hooks.
A positive `--dlo-runtime-cache-pin-limit-gib` deliberately selects this final
runtime layout even when a direct TP1 checkpoint plan is available. Raw
checkpoint mappings may require deferred loader transforms such as grouped-QKV
reordering; registering them would still leave recurrent host work. Publishing
the fully transformed layout once makes every streamed source directly
copyable.
The runtime layout identity excludes DP rank and SP rank. It includes TP world
size/rank and conservative SP implementation/world-size guards. Therefore two
@@ -110,13 +122,27 @@ independent `TP=2` engines normally create two entries: both TP0 workers map
one entry and both TP1 workers map the other. Cache sharing does not create a
process group or an inference-time collective.
At inference time, cached tensors remain pageable mmap views. DLO packs each
streamed block into one of two bounded pinned host slots before the asynchronous
H2D copy. The private ordinary-loader path instead creates model-sized pinned
storage once during initialization. The cache therefore exchanges recurrent
host-to-host copies and DDR bandwidth for lower steady-state PSS; competing TP
ranks or engines can turn staging skew into collective wait time. Treat this as
a host-capacity mode and benchmark it on the target CPU/memory topology.
With a zero registration budget, cached tensors remain pageable mmap views.
DLO packs each streamed block into one of two bounded pinned host slots before
the asynchronous H2D copy. This lowers steady-state PSS but adds recurrent
host-to-host copies; competing TP ranks or engines can turn DDR contention and
staging skew into collective wait time.
With a positive budget on CUDA, DLO instead attempts to register the complete
page-aligned cache mapping as read-only pinned host memory. Registration is
all-or-nothing: the total mapped range must fit the per-worker budget, and any
CUDA registration or verification failure rolls back partial registrations and
uses the bounded staging path. If CUDA cannot roll back a partial registration,
worker initialization aborts instead of unmapping memory that CUDA may still
own. On success, each immutable tensor view is copied directly into the
existing rotating device buffers on the H2D stream, so no private host staging
slots are allocated.
CUDA registration is process-local, but it does not clone file contents. The
same physical page-cache pages remain shared among equivalent workers; each
worker does, however, hold its own registration reference and must satisfy the
host's page-locking limits. Shutdown synchronizes outstanding GPU work,
unregisters every range, and only then closes the mmap handles.
The mmap plan skips only dedicated DiT weight sources. Other component sources,
such as a text encoder loaded through the shared diffusion loader, continue to
@@ -165,6 +191,13 @@ entry is published. Operators may remove entries or the complete cache root
only after all workers using its mmap files have stopped. The feature is
opt-in for this reason.
For direct H2D, set the registration limit above the complete page-aligned
layout size reported in the worker log. Partial registration is intentionally
unsupported because mixing direct and staged tensors would complicate buffer
lifetime and make performance topology-dependent. Registration is CUDA-only;
an insufficient budget, unsupported runtime, OS/CUDA pinning limit, or
registration error is reported and falls back to bounded staging.
## Declarative topology
Models may declare an `OffloadPlan` instead of embedding offload logic:
@@ -196,7 +229,9 @@ must enter each collective.
- Direct checkpoint mmap currently requires TP1. TP greater than one falls
back to the ordinary TP-aware loader. With the optional runtime cache,
equivalent processes at the same TP coordinate can then share that final
rank-local layout in no-AllGather mode.
rank-local layout in no-AllGather mode. A positive registration budget also
chooses the final runtime cache at TP1 so deferred checkpoint transforms are
not repeated during inference.
- HSDP plus AllGather is rejected to avoid double sharding. HSDP without
AllGather has limited end-to-end validation.
- Per-tensor online FP8 linears use the ordinary loader and can run with either
@@ -212,16 +247,19 @@ must enter each collective.
The runtime cache improves steady-state host PSS, not startup peak: every rank
loads private weights before remapping, and validation adds full content-hash
passes. Skip-load-on-cache-hit and cache eviction remain follow-up work in
passes. Direct H2D removes recurrent staging but does not change that startup
contract. Skip-load-on-cache-hit and cache eviction remain follow-up work in
[RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).
It is also not latency-neutral. In the four-L20X MiniMax-H3 smoke documented in
the L20X runtime-cache matrix in the
Zero-budget staging is also not latency-neutral. In the four-L20X MiniMax-H3
smoke documented in the L20X runtime-cache matrix in the
[feature design](../../../design/feature/offloader/distributed_layerwise_offload.md#l20x-runtime-cache-smoke-matrix),
two independent TP2 engines reduced request-time host PSS by 36.6% with the
cache, but recurrent mmap-to-pinned staging reduced combined throughput by
69.3%. HBM and GPU compute were unchanged. Use the cache only when that
measured memory/latency tradeoff is appropriate for the deployment.
69.3%. The registered path recovered near-private per-engine latency while
retaining the shared-memory benefit; a matched TP1 run improved mean wave
latency from 33.87 seconds to 4.88 seconds with byte-identical outputs. See the
feature design for the full smoke-test caveats and metrics.
See the [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
for an end-to-end example.