Master commit cf758bd2 (PR #13663, "chore(api-nodes): increase default
timeout for partner API node tasks") removed three explicit
max_poll_attempts=280 overrides from nodes_kling.py so the new 480
default in util/client.py would take effect.
The May 19 merge of master into worksplit-multigpu (ff766e5c) silently
discarded those three deletions in the 3-way resolve - nodes_kling.py
had no textual conflict but the resolution kept the pre-cf758bd2 lines.
The other seven files cf758bd2 touched were merged correctly; this
restores nodes_kling.py to match master.
Amp-Thread-ID: https://ampcode.com/threads/T-019e52b4-31ee-72cd-996b-64ecd9420e13
Co-authored-by: Amp <amp@ampcode.com>
* openapi: add enum values + FeedbackRequest schema for cloud cutover (PR E)
Adds missing cloud-runtime enum values to vendor schemas that the
cloud runtime emits but vendor declared as plain strings.
Changes:
- JobEntry.status: enum [pending, in_progress, completed, failed, cancelled]
- JobDetailResponse.status: same enum
- BillingStatus: enum [awaiting_payment_method, pending_payment, paid,
payment_failed, inactive]
- FeedbackRequest schema added (with type enum)
- /api/feedback POST: requestBody now $refs FeedbackRequest
All cloud-runtime-emitted; no impact on OSS-local semantics.
Identified via Comfy-Org/cloud's TestCutoverSafe gate (BE-1106) as
the remaining schema-level divergences after PRs A-D landed and got
synced.
* openapi: add type enum to Workspace schema (cutover follow-up)
Cloud's Workspace runtime shape includes a 'type' field with enum
[personal, team] that vendor's Workspace was missing. Cloud handlers
reference the generated ingest.WorkspaceType Go enum.
Same kind of surgical addition as JobEntry.status / BillingStatus /
JobDetailResponse.status in this PR — adds cloud-runtime field to
existing vendor schema.
Two fixes for single-GPU users on non-NVIDIA backends; multi-GPU
non-CUDA support is intentionally out of scope here (tracked separately).
1. get_all_torch_devices: add AMD/ROCm, MLU, and a generic fallback arm.
Previously the function only enumerated NVIDIA, Intel XPU, and Ascend
NPU when cpu_state==GPU; on AMD/ROCm (which exposes its GPU through
torch.cuda.*) and DirectML it fell through to an empty list. The
biggest user-visible regression: unload_all_models() iterates this
list, so it became a silent no-op on AMD/ROCm. /free, manager
unloads, and shutdown stopped releasing VRAM.
- is_amd() now shares the torch.cuda.* arm with is_nvidia(), since
ROCm reuses the CUDA API surface.
- is_mlu() gets its own arm using torch.mlu.device_count().
- A final fallback appends get_torch_device() for any GPU backend
the explicit arms miss (notably DirectML), so callers see at
least the current device and unload_all_models works.
MPS users are unaffected: cpu_state==MPS already routes to the
else branch which appends get_torch_device() returning mps.
2. main.py DynamicVRAM init: guard the comfy_aimdo branch with an
explicit is_nvidia() check.
The outer condition allows entering the DynamicVRAM init block when
the user passes --enable-dynamic-vram explicitly, bypassing the
implicit is_nvidia() gate. On non-NVIDIA backends this then runs
comfy_aimdo.control.init_devices(range(torch.cuda.device_count())),
which is comfy-aimdo-only territory and may crash at startup. Add a
leading is_nvidia() check that logs a clean warning and falls back
to the legacy ModelPatcher path.
* Revert "Add tiled VAE lane to MultiGPU Work Units"
This reverts commit 4d3d68e473.
The tiled VAE lane will land as part of a follow-up PR alongside the
UPSCALE_MODEL lane, separated from the threaded-loader fix PR (#14052)
to keep the upstream merge focused.
* Revert "Add UPSCALE_MODEL lane to MultiGPU CFG Split"
This reverts commit 74b0a826ea.
The UPSCALE_MODEL lane will land as part of a follow-up PR alongside the
tiled VAE lane, separated from the threaded-loader fix PR (#14052) to
keep the upstream merge focused.
---------
Co-authored-by: John Pollock <pollockjj@gmail.com>
* openapi: rename cloud-side response schemas to match runtime (PR D)
Follow-up to the BE-1106 stack (#14060/61/63). Cloud's Go handlers
reference response schemas by name (e.g., ingest.WorkflowResponse,
ingest.SubscribeResponse), but vendor's matching operations were
declaring those responses against differently-named vendor-side
schemas (CloudWorkflow, BillingSubscription, etc.). After the stack
landed, schemas like WorkflowResponse exist in vendor but weren't
referenced by any path, so codegen pruned the unreferenced types.
This PR:
1. Updates 34 operation $refs in cloud-runtime paths to point to
the schema names cloud's handlers expect (e.g., CloudWorkflow →
WorkflowResponse on /api/workflows/{workflow_id}).
2. Adds 12 cloud-only schemas that weren't in vendor yet but are
referenced by these renames (e.g., SubscribeResponse,
CancelSubscriptionResponse, BillingOpStatusResponse). Each
copied verbatim from Comfy-Org/cloud's hand-written ingest spec
and tagged x-runtime: [cloud] with a [cloud-only] description
prefix.
Schema renames span the same domains as the operationId renames in
PR A: billing/subscriptions (7 schemas), workflows (5), userdata (3),
jobs (2), hub (2), history (2), auth/workspace (4), and misc cloud
endpoints (9).
Convergent safety check after this lands (against cloud's
TestCutoverSafe gate, BE-1106):
Pre-PR D: 205 missing handler refs
Post-PR D: 105 missing handler refs (-49%)
Cumulative since the original 938-ref baseline: -89%
The remaining 105 are a Phase 3 follow-up (response headers,
text/plain responses, codegen-derived enum sub-types, and a small
set of inline-response-schema operations that vendor declares
inline where cloud has named-schema $refs).
* openapi: drop PR-label comment from new schemas block
PR-internal labels don't belong in committed code — future readers
won't know what 'PR D' means and the marker stops being useful the
moment this PR merges.
* openapi: rename 55 cloud-side operationIds to match runtime handlers
For the 55 operations below, vendor's operationId did not match the
name cloud's runtime handlers expect. Generated types from vendor
therefore had different names (e.g. CreateSubscription200JSONResponse)
than what cloud handlers reference (Subscribe200JSONResponse), which
blocks the post-cutover combined-spec codegen.
All 55 renames target the cloud-runtime-authoritative name. Several
of these endpoints are shared concepts (queue, settings, userdata,
object_info) that OSS local also serves — the rename aligns vendor
with the longstanding cloud handler-side convention to unblock the
shared codegen. No request/response *shape* changes in this PR; only
operationId labels.
Notable categories:
- Billing/subscriptions: 7 renames (subscribe, getBillingPlans, ...)
- Workspace + workflows: 13 renames (createWorkflow, ...)
- Hub: 3 renames
- Auth/users: 5 renames
- Shared OSS surface (settings, queue, view, userdata): 12 renames
- Misc cloud-only: 15 renames
Identified via Comfy-Org/cloud's TestCutoverSafe build-safety gate
(BE-1106), which compares handler type references against codegen
output from the combined spec.
* fix(openapi): resolve getHistory operationId collision
Spectral flagged: both /api/history (OSS local) and /api/history_v2
(cloud) had operationId 'getHistory' after the rename. Rename vendor's
/api/history to 'getPromptHistory' to disambiguate. Cloud's runtime
denies /api/history at the overlay level so combined codegen is
unaffected by this change.
* openapi: add 41 cloud-runtime schemas to components.schemas (PR B of 3) (#14061)
* openapi: add 41 cloud-runtime schemas to components.schemas (cutover prep)
Adds schemas that exist in Comfy-Org/cloud's hand-written ingest spec
but not yet in this vendored OSS spec. All tagged x-runtime: [cloud]
per the field-drift convention and prefixed with [cloud-only] in the
description.
These schemas are referenced by cloud's Go handlers via the generated
ingest.<Schema> Go type names. Codegen from the vendored spec didn't
produce those types because the schemas weren't declared here. Adding
them unblocks the post-cutover combined-spec codegen.
Schemas added (alphabetical):
AssetDownloadResponse, AssetMetadataResponse, BillingBalanceResponse,
BillingPlansResponse, BillingStatusResponse, GetUserDataResponseFull,
HistoryDetailEntry, HistoryDetailResponse, HistoryResponse,
HubLabelInfo, HubProfileSummary, HubWorkflowListResponse,
HubWorkflowStatus, HubWorkflowSummary, HubWorkflowTemplateEntry,
JobStatusResponse, JobsListResponse, LabelRef, LogsResponse, Member,
OAuthRegisterBadRequestResponse, PendingInvite, Plan, PlanAvailability,
PlanAvailabilityReason, PlanSeatSummary, PreviewPlanInfo,
PreviewSubscribeResponse, PublishedWorkflowDetail, SecretResponse,
SubscriptionDuration, SubscriptionTier, UserDataResponseFull,
ValidationError, ValidationResult, WorkflowForkedFrom, WorkflowResponse,
WorkflowVersionContentResponse, WorkspaceAPIKeyInfo, WorkspaceSummary,
WorkspaceWithRole
Identified via Comfy-Org/cloud's TestCutoverSafe build-safety gate
(BE-1106). Companion to PR #14060 (operationId renames).
* fix(openapi): add BindingErrorResponse schema
OAuthRegisterBadRequestResponse references BindingErrorResponse but
that schema wasn't in the original add. Adding it now as a cloud-only
schema matching the cloud runtime's binding-error shape (single
'message' string field).
* openapi: add missing 4xx/5xx response bodies for cloud-emitting endpoints (#14063)
Vendor declares shared endpoints (e.g. /api/queue, /api/settings,
/api/assets/*, /api/billing/*) with success responses but is missing
many of the 4xx/5xx error response bodies that Comfy-Org/cloud's
runtime actually emits. Cloud's Go handlers reference the generated
ingest.Op<StatusCode>JSONResponse types for these missing statuses,
which currently fail to resolve when codegen runs against the
vendored spec.
This PR adds 237 response entries across 117 operations, restoring
the documented error responses that cloud emits. Bodies are copied
verbatim from Comfy-Org/cloud's hand-written ingest spec
(services/ingest/openapi.yaml) and reference a new ErrorResponse
schema also added in this PR (matches cloud's {code, message} runtime
shape, tagged x-runtime: [cloud]).
ErrorResponse is intentionally separate from the existing CloudError
schema. CloudError's shape ({error}) describes one runtime; cloud
emits a different shape ({code, message}). Existing CloudError refs
in vendor are untouched; new cloud-emitting error references use
ErrorResponse.
Identified via Comfy-Org/cloud's TestCutoverSafe build-safety gate
(BE-1106). Companion to PR #14060 (operationId renames) and PR #14061
(cloud-only schema additions).
* openapi: align response declarations with implementation (5 endpoints)
- POST /api/assets/download: replace 200 with 202 + tracking-task body
(endpoint runs asynchronously and returns task_id/status/message).
- POST /api/assets/export: same 200 → 202 + tracking-task body.
- POST /api/assets/from-workflow: change 201 → 200 (handler responds 200,
not 201; no Location header emitted).
- POST /api/feedback: change 200 → 201 (creates a feedback record).
- /api/jobs and /api/jobs/{job_id}: change timestamp fields from
type: number to type: integer + format: int64. Values are Unix
milliseconds — number causes oapi-codegen to emit float64, losing
precision and producing the wrong Go type. Affected fields:
create_time, update_time, execution_start_time, execution_end_time.
Verification: each change reflects what the endpoint observably returns;
no handler changes required. Backwards-compatible for existing clients
(integer is a subset of number; status code shifts within 2xx).
* openapi: align asset download/export 202 status enum with runtime + sibling schemas
CodeRabbit caught a vocabulary mismatch: the two new 202 response schemas
declared `[pending, running, completed, failed]` while the rest of the same
spec uses `[created, running, completed, failed]` for the identical task
lifecycle (download/export progress WebSocket events, /api/tasks, TaskEntry,
TaskResponse — 4 sites total). Cloud's runtime emits `created` on initial
creation (AssetDownloadResponseStatusCreated; task.Status sourced from the
DB enum whose initial value is Created). `pending` would have introduced a
fifth, contradictory vocabulary for the same lifecycle and pushed the spec
further from the implementation it is meant to align with.
Followup tracked separately: extract a shared TaskStatus enum so all five
sites move in lockstep instead of needing per-site edits.
Introduce tiled_scale_multidim_multigpu in comfy/utils.py: a tile scheduler
that dispatches per-device tile functions through the existing
MultiGPUThreadPool and merges per-device CPU output buffers in deterministic
key order. The worker only catches BaseException at the thread boundary to
funnel errors to the main thread; bare torch.cuda.set_device and
torch.cuda.synchronize calls inside the worker fail loud if the device is
not CUDA, which is part of the primitive's contract.
Add UPSCALE_MODEL input on the MultiGPU CFG Split node and an upscale-model
descriptor deepclone helper in comfy/multigpu.py. Clones stay CPU-resident
until execute time and are returned to CPU afterward.
ImageUpscaleWithModel dispatches through tiled_scale_multidim_multigpu when
a multigpu descriptor is attached; the single-device path runs unchanged
when no clones are present.
Comfy-aimdo 0.4.4 contains a small bugfix to allow recovery of a hostbuf
after full truncation.
This pattern doesnt happen as a general rule, but does happen in the
upcoming worksplit-multigpu branch.
This was an attempt to be a fast path by ensuring the file slice was
created by the owning thread and refusing without needing ot mutex
but worksplit-multigpu doesnt work that way. Go mutex.
Shoot me for overthinking next time.
The /system_stats endpoint was returning a hardcoded single-element
devices list built from get_torch_device(), which only reflects the
primary CUDA device. On multi-GPU systems this hides the additional
devices from frontends / tooling (the API surface that enables multigpu
support discovery). Switch to iterating get_all_torch_devices(), with
the primary device kept first so existing clients reading devices[0]
keep working.
(Worksplit-multigpu-only: get_all_torch_devices is the multigpu helper
introduced on this branch; master's /system_stats remains unchanged.)
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
Two CodeRabbit findings from #7063 (#13 and #14) are deferred because
worksplit-multigpu's initial release scope is NVIDIA-only QA. Leave a
TODO at the unconditional torch.cuda.set_device call and at the
post-aggregation point so the required guards/synchronize are easy to
find when multigpu support is extended to XPU/NPU/MPS/CPU/DirectML.
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
Brings in 18 commits from master so worksplit-multigpu does not regress
fixes that landed on main since the last sync:
- #13699 Hunyuan 3D 2.1 batch-size fixes (overlap with our own backport;
conflict resolved in favor of the shape>=2 gate that binds
swap_cfg_halves once and reuses it for the output swap-back)
- #14031 ModelPatcherDynamic lora reshape / backup restore fix
- #13802 Multi-threaded model load (memory_management / pinned_memory /
model_management / aimdo plumbing)
- #12679 lanczos single-channel tensor fix
- #14010 Stable Audio 3 support
- assorted partner-node, openapi, workflow-template, and tooling updates
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
CrossAttention.kv.view and Attention.qkv_combined.view both hardcoded
batch=1 in the reshape, crashing or silently mis-shaping whenever the
actual batch dimension was greater than 1. These were fixed on master
in #13699 as part of the same patch that gated the chunk(2) swap, but
worksplit-multigpu only picked up the chunk(2) gate. Bring the two
view() fixes over so we have parity with master.
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
The previous gate (len(cond_or_uncond) == 2 and set == {0, 1}) was
intended to skip the cond/uncond swap when only one half was present
under MultiGPU CFG Split, but it was too restrictive: it also skipped
batch_size > 1 + CFG (cond_or_uncond like [0, 0, 1, 1] or [0,0,0,0,
1,1,1,1]), where chunk(2) still splits the batch cleanly into a cond
half and an uncond half and the swap is still required.
Switch to context.shape[0] >= 2, matching the parallel fix landed on
master in #13699. The swap is a permutation-invariant no-op when the
two halves don't form a CFG pair (since the output swap_cfg_halves
block immediately undoes the permutation), so the only thing the gate
actually needs to do is guard against chunk(2) on a batch of one.
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
Per review feedback on #7063. The two functions share the conds-by-hooks
accumulation, memory-fit batching, and per-chunk output aggregation; the
multigpu variant adds per-device scheduling, .to(device) placement,
per-device patcher/control lookup, and thread-pool dispatch around the
inner loop. Documenting the relationship without extracting helpers --
extraction can land after the initial worksplit-multigpu release once
both paths have settled.
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
QwenFunControlNet.pre_run stashes model.diffusion_model into extra_args,
which the control_model then uses for forward passes (img_in, txt_in,
pe_embedder, time_text_embed). With multigpu, every per-device control
clone was being pre_run with the base model on GPU0, so secondary
devices would invoke those modules with parameters on GPU0 and inputs
on their own device, raising 'Expected all tensors to be on the same
device'. Build a device -> per-device BaseModel lookup from the
patcher's additional multigpu models and pass each clone the model on
its own device. Falls back to the base model when no per-device match
is found (single-GPU path and the case where cnet.multigpu_clones lags
the patcher's clone set).
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
QwenFunControlNet.pre_run stashes the model's diffusion_model into
self.extra_args['base_model'], but ControlBase.cleanup never clears
extra_args. The diffusion_model reference therefore lingered between
sampling runs, blocking ComfyUI's model offload/eviction logic from
freeing the UNet and -- for multigpu -- holding one such reference per
per-device control clone (defeating the max_gpus pruning added in this
PR). Override cleanup to drop the entry; super().cleanup() already
recurses into multigpu_clones so each per-device clone pops its own.
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
Drop the new ignore_multigpu positional argument from prepare_state and
from the ON_PREPARE_STATE callbacks; pass the flag via model_options
instead. This restores the original 3-arg callback signature so existing
custom-node ON_PREPARE_STATE handlers keep working unchanged, while
still letting prepare_state's recursive call into multigpu_clones
short-circuit.
Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082
Co-authored-by: Amp <amp@ampcode.com>
* ModelPatcherDyanmic: purge stale vbar allocs on force cast
* ModelPatcherDynamic: restore backups before load
If doing a clean reload, mutative changes (lora application) could be
applied on-top of the already loaded weight. Restore from backup
unconditionally so that the new load is clean.
The job_ids query parameter on GET /api/assets is tagged x-runtime:
[cloud] and only exists for cloud's variant of this endpoint. Cloud
removed all consumers and the cloud-side handler/codegen/tests in
Comfy-Org/cloud#3778. With cloud no longer accepting this parameter,
the [cloud-only] documentation here is wrong — drop it so the daily
sync to cloud/services/ingest/vendor/openapi.yaml propagates the
removal.
The operation at POST /api/assets/import was defined as `importAssets`
with a URL-list body shape, but no runtime actually serves that
operation at this path. The cloud runtime serves a different operation
here — `importPublishedAssets` — which imports published-workflow
assets into the caller's library by ID, not by URL.
Cloud's URL-based asset ingestion lives at separate paths
(POST /assets/download + GET /assets/remote-metadata) tracked
elsewhere; nothing in this PR affects that work.
Changes:
- Replace the operation at POST /api/assets/import with
`importPublishedAssets`, taking ImportPublishedAssetsRequest
(published_asset_ids + optional share_id) and returning
ImportPublishedAssetsResponse (list of AssetInfo).
- Remove the unused AssetImportRequest component schema (no other
references in the spec).
- Operation and schemas tagged x-runtime: [cloud] with [cloud-only]
description prefix, matching the existing convention for
cloud-runtime-only operations elsewhere in the spec.
Spectral lint passes (0 errors); the two hint-level findings on
the spec are pre-existing and unrelated.
No FE consumer references AssetImportRequest today; this is a pure
spec correction to match what the cloud runtime actually serves.
Add the OAuth 2.1 authorization flow and RFC 7591 Dynamic Client
Registration endpoints to the shared spec, alongside the existing
auth-tagged operations (/api/auth/session, /api/auth/token,
/.well-known/jwks.json). All tagged x-runtime: [cloud] with a
[cloud-only] description prefix, following the established
convention for cloud-runtime-only operations.
Endpoints:
- GET /.well-known/oauth-authorization-server (RFC 8414 metadata)
- GET /.well-known/oauth-protected-resource (RFC 9728 metadata)
- GET /oauth/authorize (consent challenge)
- POST /oauth/authorize (consent submission)
- POST /oauth/token (RFC 6749 §3.2)
- POST /oauth/register (RFC 7591 §3.1 DCR)
Component schemas added:
- OAuthAuthorizationServerMetadata
- OAuthProtectedResourceMetadata
- OAuthConsentChallenge, OAuthConsentChallengeWorkspace
- OAuthAuthorizeRedirectResponse
- OAuthTokenResponse, OAuthTokenError
- OAuthRegisterRequest, OAuthRegisterResponse, OAuthRegisterError
These endpoints are implemented in the cloud runtime today and
are called by browser frontends rendering the consent UI and by
MCP-spec-compliant clients (Claude Desktop, Cursor, etc.) doing
auto-discovery + self-registration. Documenting them in the
shared spec lets the cloud frontend generate types directly from
this spec instead of maintaining a parallel definition.
Spectral lints clean (0 errors). The hint-level findings on
OAuthTokenError / OAuthRegisterError ("standard error schema")
match the same hint on CloudError — these are protocol-specific
RFC-shaped errors, not generic application errors.
Two doc-only changes addressing minor CodeRabbit findings on PR #7063:
* cli_args.py: clarify --cuda-device help text to document the required comma-separated format ('0' or '0,1'), matching how the value is consumed by CUDA_VISIBLE_DEVICES in main.py.
* nodes_multigpu.py: add a docstring NOTE on the (currently unregistered) MultiGPUOptionsNode explaining that its relative_speed input is plumbed through to model_options['multigpu_options'] but is not yet consulted by the cond scheduler, which still uses uniform round-robin via next_available_device(). Wire relative_speed into the scheduler before re-enabling the node.
Amp-Thread-ID: https://ampcode.com/threads/T-019e43b8-8258-70fd-ab3a-53e4c97f85d5
Co-authored-by: Amp <amp@ampcode.com>
The multigpu cond-batching loop called model.memory_required(input_shape) without conditioning shapes, while the single-GPU path at line 279 passes cond_shapes. Large conditioning tensors (e.g. video prompts, control inputs) were therefore under-counted, risking OOM at runtime when the chosen batch size was too large. Match the single-GPU pattern by building cond_shapes from each batched cond's conditioning dict and passing it to memory_required.
Amp-Thread-ID: https://ampcode.com/threads/T-019e43b8-8258-70fd-ab3a-53e4c97f85d5
Co-authored-by: Amp <amp@ampcode.com>
* model_management: disable non-dynamic smart memory
Disable smart memory outright for non dynamic models.
This is a minor step towards deprecation of --disable-dynamic-vram
and the legacy ModelPatcher.
This is needed for estimate-free model development, where new models
can opt-out of supplying a memory estimate and not have to worry
about hard VRAM allocations due to legacy non-dynamic model patchers
This is also a general stability increase for a lot of stray use cases
where estimates may still be off and going forward we are not going
to accurately maintain such estimates.
* pinned_memory: implement with aimdo growable buffer
Use a single growable buffer so we can do threaded pre-warming on
pinned memory.
* mm: use aimdo to do transfer from disk to pin
Aimdo implements a faster threaded loader.
* Add stream host pin buffer for AIMDO casts
Introduce per-offload-stream HostBuffer reuse for pinned staging,
include it in cast buffer reset synchronization.
Defer actual casts that go via this pin path to a separate pass
such that the buffer can be allocated monolithically (to avoid
cudaHostRegister thrash).
* remove old pin path
* Implement JIT pinned memory pressure
Replace the predictive pin pressure mechanism with JIT PIN memory
pressure.
* LowVRAMPatch: change to two-phase visit
* lora: re-implement as inplace swiss-army-knife operation
* prepare for multiple pin sets
* implement pinned loras
* requirements: comfy-aimdo 0.4.0
* ops: remove unused arg
This was defeatured in aimdo iteration
* ops: sync the CPU with only the offload stream activity
This was syncing with the offload stream which itself is synced with the
compute stream, so this was syncing CPU with compute transitively. Define
the event to sync it more gently.
* pins: implement freeing intermediate for pinned memory
Pinning is more important than inactive intermediates and the stream
pin buffer is more important than even active intermediates.
* execution: implement pin eviction on RAM presure
Add back proper pin freeing on RAM pressure
* implement pin registration swaps
Uncap the windows pins from 50% by extending the pool and have a pressure
mechanism to move the pin reservations om demand.
This unfortunately implies a GPU sync to do the freeing so significant
hysterisis needs to be added to consolidate these pressure events.
* cli_args/execution: Implement lower background cache-ram threshold
Limit the amount of RAM background intermediates can use, so that
switching workflows doesn't degrade performance too much.
* make default
* bump aimdo
* model-patcher: force-cast tiny weights
Flux 2 gets crazy stalls due to a mix of tiny and giant weights
creating lopsided steam buffer rotations which creates stalls.
* ops: refactor in prep for chunking
* mm: delegate pin-on-the-way to aimdo
Aimdo is able to chunk and slice this on the way for better CPU->GPU
overlap. The main advantage is the ability to shorten the bus contention
window between previous weight transfer and the next weights vbar
fault.
* bump aimdo
* pinning updates
* specify hostbuf max allocation size
There a signs of virtual memory exhaustion on some linux systems when
throwing 128GB for every little piece. Pass the actual to save aimdo
from over-estimates
* tests: update execution tests for caching
The default caching changed to ram-cache so update these tests
accordingly.
Remove the LRU 0 test as this also falls through to RAM cache.
create_multigpu_deepclones cloned the existing 'multigpu' additional_models list verbatim and never pruned entries beyond limit_extra_devices. If a workflow was previously prepared for more GPUs, reducing max_gpus would leave stale clones attached and eligible for later scheduling. Replace the TODO block with a real prune that keeps only clones whose load_device is either the model's load_device or in limit_extra_devices, and re-match clones if anything was removed.
Amp-Thread-ID: https://ampcode.com/threads/T-019e43b8-8258-70fd-ab3a-53e4c97f85d5
Co-authored-by: Amp <amp@ampcode.com>
