fix(recipes): address BCM overlay gaps from H200 NVL validation#1089
Merged
Conversation
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Enterprise Run ID: 📒 Files selected for processing (3)
📝 WalkthroughWalkthroughThis PR updates Kubernetes recipe configurations: (1) adds documentation in nvidia-dra-driver-gpu values clarifying why Estimated code review effort🎯 2 (Simple) | ⏱️ ~8 minutes Possibly related issues
Possibly related PRs
Suggested labels
Suggested reviewers
🚥 Pre-merge checks | ✅ 4✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@recipes/components/nvidia-dra-driver-gpu/values.yaml`:
- Around line 69-78: This change updates component values (see
recipes/components/nvidia-dra-driver-gpu/values.yaml) but did not include the
regenerated BOM docs; run the make target "make bom-docs" locally to regenerate
docs/user/container-images.md and add/commit the updated
docs/user/container-images.md to this PR so the BOM reflects the values.yaml
change.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: ASSERTIVE
Plan: Enterprise
Run ID: 1fda92e1-e76d-445a-9a1e-e5e981fd3338
📒 Files selected for processing (3)
recipes/components/nvidia-dra-driver-gpu/values.yamlrecipes/overlays/bcm-training.yamlrecipes/overlays/bcm.yaml
yuanchen8911
force-pushed
the
fix/recipe-gaps-1086
branch
from
fd65476 to
7753de2
Compare
4 tasks
|
Actionable comments posted: 0 |
yuanchen8911
changed the title
15 tasks
yuanchen8911
changed the title
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. Resolves to the same hydrated values as h100 today via
shared base components (both are Hopper, same R570/R580 driver line,
NVML auto-detects everything); no new overlay file is needed.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0;
H100 overlay chain produces correct hydrated values. End-to-end:
`aicr recipe --accelerator h200 --service bcm --os ubuntu --intent training`
succeeds and produces a recipe with `criteria.accelerator: h200`.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
force-pushed
the
fix/recipe-gaps-1086
branch
from
7753de2 to
bccf6ad
Compare
|
Actionable comments posted: 0 |
- Document why nvidia-dra-driver-gpu controller and kubeletPlugin priorityClassName are explicitly neutralized (PSA / PriorityClass admission constraints AICR cannot assume cluster-wide). Notes the eviction-under-node-pressure trade-off so operators can re-pin via their own overlay if needed. - Mirror controller.tolerations onto kubeletPlugin.tolerations in bcm.yaml so DRA's kubelet plugin DaemonSet schedules on small BCM deployments that combine control-plane and worker roles on the same node. - Enable GPUDirect Storage (gds.enabled: true) in bcm-training.yaml. BCM-provisioned nodes typically ship NVIDIA-validated NVMe + ConnectX hardware where GDS delivers a meaningful training I/O perf win (most pronounced on H200 NVL given its 141GB HBM3e per device). Benign on nodes without compatible hardware. Surfaced during cluster-side validation of the recently-merged feat/bcm-service-type work on a real H200 NVL test cluster. Addresses 3 of 4 checkboxes in NVIDIA#1086; H200 criteria registration is the larger 4th item and will land in a separate PR per the umbrella issue.
yuanchen8911
force-pushed
the
fix/recipe-gaps-1086
branch
from
bccf6ad to
c93c5d4
Compare
|
Actionable comments posted: 0 |
mchmarny
approved these changes
May 28, 2026
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
yuanchen8911
added a commit
to yuanchen8911/aicr
that referenced
this pull request
May 28, 2026
Adds CriteriaAcceleratorH200 to the criteria registry so users can
pass `--accelerator h200` and have the recipe metadata reflect the
hardware. H200 is the same Hopper generation as H100 (same R570/R580
driver line, same gpu-operator support floor), so its deployment-phase
floor mirrors H100's.
Updates every surface that enumerates accelerator types per the
"Adding a new enum value" audit rule in .claude/CLAUDE.md:
- pkg/recipe/criteria.go: new const, parse case, AllCriteriaAcceleratorTypes
- pkg/recipe/criteria_test.go: parse + Get tests cover h200
- pkg/recipe/criteria_registry_{,parse_}test.go: swap h200 (now built-in)
for mi300x as the extensibility-test example value
- api/aicr/v1/server.yaml: all 5 enum blocks
- .github/ISSUE_TEMPLATE/bug_report.yml: GPU type dropdown entries
- docs/{README.md,user/cli-reference.md,user/api-reference.md,
contributor/api-server.md,contributor/cli.md,contributor/data.md,
contributor/validations.md,contributor/api-server-extending.md}
- pkg/{api,cli/recipe.go,fingerprint/{doc.go,types.go},recipe/doc.go,
server/doc.go}: godoc enumerations
Wires up the two consumer paths the enum alone left incomplete:
- pkg/fingerprint/gpu_sku.go: add the H200 ProductName pattern so the
snapshot -> fingerprint -> recipe path resolves real H200 hardware
("NVIDIA H200 NVL", "...141GB HBM3e") to h200 instead of unknown-sku
- recipes/overlays/h200-any.yaml: new criteria-wildcard overlay mirroring
h100-any.yaml (4 standard deployment checks + gpu-operator >= v24.6.0)
so an `--accelerator h200` recipe inherits the same deployment-phase
floor as H100 rather than landing on bare base
- .claude/skills/analyzing-snapshots/SKILL.md: add h200 to the
model->accelerator mapping and valid-values tables
Validated against a real H200 NVL cluster: GFD / DRA correctly identify
the device as "NVIDIA H200 NVL", 141GB HBM3e, Hopper, compute 9.0.
End-to-end: `aicr recipe --accelerator h200 --service bcm --os ubuntu
--intent training` resolves the identical deployment floor as the h100
equivalent and produces a recipe with criteria.accelerator: h200.
Addresses checkbox 4 of NVIDIA#1086 (the H200 registration item carved out
from PR NVIDIA#1089).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three small recipe-layer edits surfaced during cluster-side validation of the recently-merged
feat/bcm-service-typework on a real H200 NVL test cluster: document why DRApriorityClassNameis neutralized, mirror BCM control-plane tolerations onto DRA'skubeletPlugin, and enable GPUDirect Storage for BCM training.Motivation / Context
Fixes: part of #1086 (3 of 4 checkboxes; H200 criteria registration deferred to a separate PR per the umbrella issue)
Related: #1087 (integration test for
nvidiaDriverRoot/hostPaths.driverInstallDirlockstep)Cluster-side validation context:
node007aicr validatedeployment+conformance phases pass end-to-endType of Change
Component(s) Affected
pkg/recipe)Implementation Notes
recipes/components/nvidia-dra-driver-gpu/values.yaml— Document the existingpriorityClassName: ""neutralization rather than restoring the upstream chart default (system-node-critical). The override has been in place sincechore: init repo(Jan 2026) with no historical PR rationale; the most plausible reason is PriorityClass admission restrictions (PSA, ResourceQuota, PriorityClassPolicy) that AICR cannot assume across all supported services. Restoring the chart default would be a behavior change without a clear test signal, so the conservative path is documentation + aTODO(#1086)for revisit. Operators who need DRA pods to survive node-pressure eviction can re-pin via their own overlay.recipes/overlays/bcm.yaml— Mirrorcontroller.tolerationsontokubeletPlugin.tolerations. Followup investigation revealed the AICR bundler appends a blanket{operator: Exists}toleration to both paths viaregistry.yamlnodeScheduling.{system,accelerated}.tolerationPathsentries (defaults sourced frompkg/snapshotter/agent.go DefaultTolerations). The blanket subsumes the specific master+control-plane entries, so the mirror is functionally a no-op in default mode — its purpose is symmetry with the existing controller block and override-resilience when a user passes--system-node-tolerationto drop the blanket. An inline comment documents this relationship so future overlay editors aren't misled. Filed separately as an AICR design question whether the tolerate-all default should be tightened per-service.recipes/overlays/bcm-training.yaml— Enablegds.enabled: truefor thegpu-operatorcomponentRef. BCM-provisioned nodes typically ship NVIDIA-validated NVMe + ConnectX hardware where GDS is a meaningful training I/O perf win (most pronounced on H200 NVL given 141GB HBM3e per device). On nodes without compatible hardware thenvidia-fsDaemonSet is benign — it logs a warning and stays inert.Testing
YAML-only changes; no Go source modified. Per
CLAUDE.local.mdscoped verification policy, ran the checks that match the change surface:yamllint -c .yamllint.yaml recipes/components/nvidia-dra-driver-gpu/values.yaml \ recipes/overlays/bcm.yaml \ recipes/overlays/bcm-training.yaml # Clean. go test -count=1 -timeout 180s ./pkg/recipe/... # ok github.com/NVIDIA/aicr/pkg/recipe 0.864s # ok github.com/NVIDIA/aicr/pkg/recipe/oskind 0.713s go test -count=1 -timeout 240s ./pkg/bundler/... # 15 packages, all ok (recipe values flow through the bundler) # End-to-end recipe + bundle against the BCM overlay chain: aicr recipe --service bcm --accelerator h100 --os ubuntu --intent training -o /tmp/recipe.yaml aicr bundle -r /tmp/recipe.yaml --deployer helmfile -o /tmp/bundle # Generated DRA values show: # - controller.tolerations: [master, control-plane, {operator: Exists}] # - kubeletPlugin.tolerations: [master, control-plane, {operator: Exists}] # - gds.enabled: true on gpu-operatorSkipping full
make qualifybecause the change is YAML-only — no Go test, e2e, lint, or scan target can regress from these edits. The scoped recipe + bundler tests cover the consumers that parse these files.Risk Assessment
priorityClassName(documentation only); the kubeletPlugin tolerations mirror is a no-op in default mode (bundler blanket subsumes it); the BCM GDS enable is additive and benign on hardware that doesn't use them.Rollout notes: None. The GDS daemon is benign on nodes without compatible NVMe/NIC hardware (it logs a warning and stays inert).
Checklist
./pkg/recipe/...and./pkg/bundler/...with-count=1)yamllinton changed files)git commit -S)