Breadway/breadram
1
0
Fork 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

Read and report integrated GPU shared memory (GTT) #35

Merged
Breadway merged 4 commits from feature/ram-67-igpu-gtt into main 2026-10-12 00:39:03 +08:00
Conversation 1 Commits 4 Files changed 13 +446 -20
Breadway commented 2026-10-11 23:52:12 +08:00
Owner
Copy link

The first piece of integrated GPU (iGPU) support: reading and reporting GTT shared memory.

Part of RAM-67. On an iGPU the device memory is a carve-out of host RAM, so it must be counted once and reported separately from host available memory and from dedicated VRAM.

What this adds

  • breadram-core/src/telemetry/igpu.rs: reads mem_info_gtt_used and mem_info_gtt_total under /sys/class/drm/card<N>/device, bounded to a fixed number of cards. A card with neither counter is not an integrated GPU.
  • The memory breakdown gains gtt_used_bytes and gtt_total_bytes (protocol minor 7, both defaulted), summed across cards, shown by breadram diagnostics (human and JSON), and surfaced as the gpu_gtt capability: available when a used count is read, degraded when only a total is, unavailable with a reason when the drm class is unreadable or there is no iGPU.
  • Missing is never zero: an unreadable or absent counter reports unknown with a reason.

Verification

Run on this branch off main:

  • cargo fmt --all --check
  • cargo clippy --workspace --all-targets --locked -- -D warnings
  • cargo build --workspace --locked
  • cargo test --workspace --locked

All pass. Tests cover a two-card sum, a dedicated card that is not an iGPU, an unreadable drm class, and a used count with no total.

Still to come on RAM-67

Attributing an iGPU-backed model's footprint through the inference adapters so an otherwise idle workload is not released while it holds a loaded iGPU model, counting the GTT carve-out once when it overlaps cgroup accounting, and exercising the reading against a real iGPU host.

The first piece of integrated GPU (iGPU) support: reading and reporting GTT shared memory. Part of RAM-67. On an iGPU the device memory is a carve-out of host RAM, so it must be counted once and reported separately from host available memory and from dedicated VRAM. ## What this adds - `breadram-core/src/telemetry/igpu.rs`: reads `mem_info_gtt_used` and `mem_info_gtt_total` under `/sys/class/drm/card<N>/device`, bounded to a fixed number of cards. A card with neither counter is not an integrated GPU. - The memory breakdown gains `gtt_used_bytes` and `gtt_total_bytes` (protocol minor 7, both defaulted), summed across cards, shown by `breadram diagnostics` (human and JSON), and surfaced as the `gpu_gtt` capability: available when a used count is read, degraded when only a total is, unavailable with a reason when the drm class is unreadable or there is no iGPU. - Missing is never zero: an unreadable or absent counter reports unknown with a reason. ## Verification Run on this branch off main: - cargo fmt --all --check - cargo clippy --workspace --all-targets --locked -- -D warnings - cargo build --workspace --locked - cargo test --workspace --locked All pass. Tests cover a two-card sum, a dedicated card that is not an iGPU, an unreadable drm class, and a used count with no total. ## Still to come on RAM-67 Attributing an iGPU-backed model's footprint through the inference adapters so an otherwise idle workload is not released while it holds a loaded iGPU model, counting the GTT carve-out once when it overlaps cgroup accounting, and exercising the reading against a real iGPU host.
feat: read and report integrated GPU shared memory (GTT)
All checks were successful
CPU checks / check (push) Successful in 55s
Details
ba0cd020e8 
First piece of RAM-67. On an iGPU the device memory is a carve-out of host RAM,
so it is reported separately and never folded into host available memory.

- telemetry::igpu reads mem_info_gtt_used and mem_info_gtt_total under the drm
  class, bounded to a fixed number of cards, and treats a card with neither
  counter as not an integrated GPU.
- The memory breakdown gains gtt_used_bytes and gtt_total_bytes (protocol minor
  7, both defaulted), shown by breadram diagnostics and surfaced as the gpu_gtt
  capability: available when a used count is read, degraded when only a total is,
  and unavailable with a reason when the drm class is unreadable or there is no
  iGPU.
- Missing is never zero: an unreadable or absent counter reports unknown.
- Tests cover a two-card sum, a dedicated card that is not an iGPU, an
  unreadable drm class, and a used count with no total.
- docs/IGPU.md records the reading and what remains (inference attribution,
  counting the carve-out once).
docs: record the live iGPU GTT reading on an AMD host
All checks were successful
CPU checks / check (push) Successful in 1m14s
Details
30ce197d9f 
Verified on this host's amdgpu integrated GPU (card1): breadramd in monitor mode
plus breadram diagnostics reported GTT used 7.5 GiB of 30.6 GiB and the gpu_gtt
capability available, with the shared-host-memory note. i915/xe hosts remain to
be exercised.
feat: attribute an iGPU model's device memory to host RAM
All checks were successful
CPU checks / check (push) Successful in 56s
Details
8d8bb87d30 
On an integrated GPU the model's reported device memory is a carve-out of host
RAM, so counting it as separate device memory would hide where it really lives.

- ProviderEndpoint gains an igpu flag (default false), and the Ollama adapter
  gains with_shared_device_memory. When set, a loaded model's whole size is host
  memory and there is no separate device figure; otherwise the two stay separate.
- The registry passes the flag through and records a note saying how the endpoint
  was attributed.
- A test asserts the iGPU attribution against the two-model fixture.
test: keep integrated GPU memory out of host available memory
All checks were successful
CPU checks / check (push) Successful in 57s
Details
a9de8f95e8 
States and tests the count-once invariant: GTT is reported in its own field and
is never folded into host available memory, because on an iGPU it is a carve-out
of the same RAM. The daemon test asserts the breakdown copies the GTT figure and
leaves ram_available_bytes exactly as the kernel reported it.
Breadway force-pushed feature/ram-67-igpu-gtt from a9de8f95e8
All checks were successful
CPU checks / check (push) Successful in 57s
Details
to 2ff1108e87
All checks were successful
CPU checks / check (push) Successful in 57s
Details
2026-10-12 00:26:08 +08:00 Compare
Breadway deleted branch feature/ram-67-igpu-gtt 2026-10-12 00:39:03 +08:00
Breadway commented 2026-10-12 00:39:20 +08:00
Author
Owner
Copy link

Handled as a duplicate. This is the auto-imported PR card for the same PR that RAM-67 tracks: #35. That PR has now been reviewed and squashed onto main as b38f03b (branch feature/ram-67-igpu-gtt deleted), so the work this card represents has landed under RAM-67. No separate delivery here. Closing it out.

**Handled as a duplicate.** This is the auto-imported PR card for the same PR that RAM-67 tracks: https://git.breadway.dev/Breadway/breadram/pulls/35. That PR has now been reviewed and squashed onto main as b38f03b (branch feature/ram-67-igpu-gtt deleted), so the work this card represents has landed under RAM-67. No separate delivery here. Closing it out.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
Breadway/breadram!35
No description provided.