Gemma-4 PLE/MoE VLM support by arielge · Pull Request #495 · torch-spyre/hf-adapters

arielge · 2026-09-08T15:45:18Z

Expand Gemma 4 multimodal support to cover encoder-free, PLE/KV-sharing, and MoE VLM variants, including E2B and 26B-A4B.
Add Spyre execution for the full Gemma 4 vision transformer, preserving two-axis RoPE, head/MLP padding, clipping, attention masks, and multimodal PLE/MoE semantics.
Refactor the dense/PLE text decoder and vision tower to use parameter-explicit compiled executors shared across structurally identical layers, substantially reducing cold-start compilation.
Add registry coverage, generic VLM correctness validation, and architecture documentation for the newly supported checkpoints.

spyre-ci · 2026-09-09T13:58:38Z

⚙️ Auto-merge enabled — dispatched /spyre-test for 980a67a2 because the required Spyre Test gate had not run on this commit.

_{Automated: enabling auto-merge requests the gate, so the PR is not left waiting on a check nobody asked for. Re-enabling auto-merge on this same commit will not dispatch again.}

spyre-ci · 2026-09-09T13:59:01Z

✅ spyre-test: success

Triggered by: /spyre-test comment

Plan (build waves + dependencies, per arch)

amd64

flowchart LR
  subgraph Lamd64_0["amd64 L0 · 1 parallel"]
    n_amd64_torch_spyre_torch_spyre_dev["torch-spyre/torch-spyre-dev 🟢<br/>image · e4b36ae9b896"]
  end
  subgraph Lamd64_1["amd64 L1 · 1 parallel"]
    n_amd64_hf_adapters_hf_adapters_dev["hf-adapters/hf-adapters-dev 🟢<br/>image · 7727c6fe8d47"]
  end
  subgraph Lamd64_2["amd64 L2 · 1 parallel"]
    n_amd64_spyre_inference_spyre_inference_dev["spyre-inference/spyre-inference-dev 🔴<br/>image · e28b30bdf999"]
  end
  n_amd64_torch_spyre_torch_spyre_dev --> n_amd64_hf_adapters_hf_adapters_dev
  n_amd64_hf_adapters_hf_adapters_dev --> n_amd64_spyre_inference_spyre_inference_dev
  classDef sPending fill:#eceff1,stroke:#90a4ae,color:#37474f
  classDef sBuilding fill:#fff8e1,stroke:#f9a825,color:#5d4037,stroke-width:2px
  classDef sOk fill:#e8f5e9,stroke:#43a047,color:#1b5e20
  classDef sReused fill:#e3f2fd,stroke:#1e88e5,color:#0d47a1
  classDef sFailed fill:#ffebee,stroke:#e53935,color:#b71c1c,stroke-width:2px
  classDef sDropped fill:#f5f5f5,stroke:#bdbdbd,color:#9e9e9e
  class n_amd64_torch_spyre_torch_spyre_dev sOk;
  class n_amd64_hf_adapters_hf_adapters_dev sOk;
  class n_amd64_spyre_inference_spyre_inference_dev sOk;

✅ orch trigger-pr-validation — green · arches amd64 · fp amd64=03df935c

level	component	arch	build	smoke	unit	integration	trunk	regression	perf
L0	torch-spyre/torch-spyre-dev	amd64	✅ ok	⏳	·	🟢 gha	·	·	·
L1	hf-adapters/hf-adapters-dev	amd64	✅ ok	🟢	·	🟢 gha	·	·	·
L2	spyre-inference/spyre-inference-dev	amd64	✅ ok	🟢	·	🟡 gha	·	·	·

GHA test runs:

[hf-adapters/amd64] integration: SUCCESS https://github.com/torch-spyre/hf-adapters/actions/runs/34361799879
[torch-spyre/amd64] integration: SUCCESS https://github.com/torch-spyre/torch-spyre/actions/runs/34361795288
[spyre-inference/amd64] integration: FAILURE https://github.com/torch-spyre/spyre-inference/actions/runs/34362314110

✅ safe to merge

Build: built 3

Tests: passed 5 · blocking 0 · advisory 0 · infra/inconclusive 0 · no signal 0

Failures by kind

ℹ️ informational · spyre-inference/amd64 gha:integration: FAILURE

Before merging, consider:

spyre-inference/amd64 gha:integration failed with gating: false — informational only, no merge impact.

spyre-ci · 2026-09-09T14:09:54Z

GHA test runs:

[torch-spyre/x86_64] integration: DISPATCHED https://github.com/torch-spyre/torch-spyre/actions/runs/34361795288

spyre-ci · 2026-09-09T14:09:56Z

GHA test runs:

[hf-adapters/x86_64] integration: DISPATCHED https://github.com/torch-spyre/hf-adapters/actions/runs/34361799879

spyre-ci · 2026-09-09T14:14:38Z

GHA test runs:

[spyre-inference/x86_64] integration: DISPATCHED https://github.com/torch-spyre/spyre-inference/actions/runs/34362314110

arielge · 2026-09-09T15:22:41Z

FYI @ani300, particularly regarding the compiled executor sharing logic

BenjSz

LGTM

spyre-ci · 2026-09-10T06:09:26Z

❌ merge-queue-integration: failure

Plan (build waves + dependencies, per arch)

amd64

flowchart LR
  subgraph Lamd64_0["amd64 L0 · 1 parallel"]
    n_amd64_torch_spyre_torch_spyre_dev["torch-spyre/torch-spyre-dev 🟢<br/>image · 21033cfc086e"]
  end
  subgraph Lamd64_1["amd64 L1 · 1 parallel"]
    n_amd64_hf_adapters_hf_adapters_dev["hf-adapters/hf-adapters-dev 🔴<br/>image · 7c49db73bf46"]
  end
  subgraph Lamd64_2["amd64 L2 · 1 parallel"]
    n_amd64_spyre_inference_spyre_inference_dev["spyre-inference/spyre-inference-dev 🟢<br/>image · c9e9ffea3cb6"]
  end
  n_amd64_torch_spyre_torch_spyre_dev --> n_amd64_hf_adapters_hf_adapters_dev
  n_amd64_hf_adapters_hf_adapters_dev --> n_amd64_spyre_inference_spyre_inference_dev
  classDef sPending fill:#eceff1,stroke:#90a4ae,color:#37474f
  classDef sBuilding fill:#fff8e1,stroke:#f9a825,color:#5d4037,stroke-width:2px
  classDef sOk fill:#e8f5e9,stroke:#43a047,color:#1b5e20
  classDef sReused fill:#e3f2fd,stroke:#1e88e5,color:#0d47a1
  classDef sFailed fill:#ffebee,stroke:#e53935,color:#b71c1c,stroke-width:2px
  classDef sDropped fill:#f5f5f5,stroke:#bdbdbd,color:#9e9e9e
  class n_amd64_torch_spyre_torch_spyre_dev sOk;
  class n_amd64_hf_adapters_hf_adapters_dev sOk;
  class n_amd64_spyre_inference_spyre_inference_dev sOk;

✅ orch trigger-pr-validation — green · arches amd64 · fp amd64=de20bf96

level	component	arch	build	smoke	unit	integration	trunk	regression	perf
L0	torch-spyre/torch-spyre-dev	amd64	✅ ok	⏳	·	🟢 gha	·	·	·
L1	hf-adapters/hf-adapters-dev	amd64	✅ ok	🟢	·	·	·	🟡 gha	·
L2	spyre-inference/spyre-inference-dev	amd64	✅ ok	🟢	·	🟢 gha	·	·	·

GHA test runs:

[spyre-inference/amd64] integration: SUCCESS https://github.com/torch-spyre/spyre-inference/actions/runs/34445225574
[torch-spyre/amd64] integration: SUCCESS https://github.com/torch-spyre/torch-spyre/actions/runs/34444526659
[hf-adapters/amd64] regression: FAILURE https://github.com/torch-spyre/hf-adapters/actions/runs/34444624409

⚠️ advisory failures only — mergeable, see below

Build: built 3

Tests: passed 4 · blocking 0 · advisory 2 · infra/inconclusive 0 · no signal 0

Failures by kind

⚠️ advisory (does not block) · hf-adapters/amd64 regression: UNSTABLE
⚠️ advisory (does not block) · hf-adapters/amd64 gha:regression: FAILURE

Before merging, consider:

hf-adapters/amd64 gha:regression, hf-adapters/amd64 regression failed with gating: "unstable" — advisory. It set the build UNSTABLE but does not block the merge. Worth a look, not a stop.

spyre-ci · 2026-09-10T06:17:07Z

GHA test runs:

[torch-spyre/x86_64] integration: DISPATCHED https://github.com/torch-spyre/torch-spyre/actions/runs/34444526659

spyre-ci · 2026-09-10T06:18:28Z

GHA test runs:

[hf-adapters/x86_64] regression: DISPATCHED https://github.com/torch-spyre/hf-adapters/actions/runs/34444624409

spyre-ci · 2026-09-10T06:26:46Z

GHA test runs:

[spyre-inference/x86_64] integration: DISPATCHED https://github.com/torch-spyre/spyre-inference/actions/runs/34445225574

spyre-ci · 2026-09-10T07:25:15Z

❌ merge-queue-integration: failure

Plan (build waves + dependencies, per arch)

amd64

flowchart LR
  subgraph Lamd64_0["amd64 L0 · 1 parallel"]
    n_amd64_torch_spyre_torch_spyre_dev["torch-spyre/torch-spyre-dev 🟢<br/>image · 21033cfc086e"]
  end
  subgraph Lamd64_1["amd64 L1 · 1 parallel"]
    n_amd64_hf_adapters_hf_adapters_dev["hf-adapters/hf-adapters-dev 🔴<br/>image · 426f303af34e"]
  end
  subgraph Lamd64_2["amd64 L2 · 1 parallel"]
    n_amd64_spyre_inference_spyre_inference_dev["spyre-inference/spyre-inference-dev 🟢<br/>image · a0d9eab46312"]
  end
  n_amd64_torch_spyre_torch_spyre_dev --> n_amd64_hf_adapters_hf_adapters_dev
  n_amd64_hf_adapters_hf_adapters_dev --> n_amd64_spyre_inference_spyre_inference_dev
  classDef sPending fill:#eceff1,stroke:#90a4ae,color:#37474f
  classDef sBuilding fill:#fff8e1,stroke:#f9a825,color:#5d4037,stroke-width:2px
  classDef sOk fill:#e8f5e9,stroke:#43a047,color:#1b5e20
  classDef sReused fill:#e3f2fd,stroke:#1e88e5,color:#0d47a1
  classDef sFailed fill:#ffebee,stroke:#e53935,color:#b71c1c,stroke-width:2px
  classDef sDropped fill:#f5f5f5,stroke:#bdbdbd,color:#9e9e9e
  class n_amd64_torch_spyre_torch_spyre_dev sOk;
  class n_amd64_hf_adapters_hf_adapters_dev sOk;
  class n_amd64_spyre_inference_spyre_inference_dev sOk;

✅ orch trigger-pr-validation — green · arches amd64 · fp amd64=51b01948

level	component	arch	build	smoke	unit	integration	trunk	regression	perf
L0	torch-spyre/torch-spyre-dev	amd64	✅ ok	⏳	·	🟢 gha	·	·	·
L1	hf-adapters/hf-adapters-dev	amd64	✅ ok	🟢	·	·	·	🟡 gha	·
L2	spyre-inference/spyre-inference-dev	amd64	✅ ok	🟢	·	🟢 gha	·	·	·

GHA test runs:

[torch-spyre/amd64] integration: SUCCESS https://github.com/torch-spyre/torch-spyre/actions/runs/34450495476
[spyre-inference/amd64] integration: SUCCESS https://github.com/torch-spyre/spyre-inference/actions/runs/34451056047
[hf-adapters/amd64] regression: FAILURE https://github.com/torch-spyre/hf-adapters/actions/runs/34450674627

⚠️ advisory failures only — mergeable, see below

Build: built 3

Tests: passed 4 · blocking 0 · advisory 2 · infra/inconclusive 0 · no signal 0

Failures by kind

⚠️ advisory (does not block) · hf-adapters/amd64 regression: UNSTABLE
⚠️ advisory (does not block) · hf-adapters/amd64 gha:regression: FAILURE

Before merging, consider:

hf-adapters/amd64 gha:regression, hf-adapters/amd64 regression failed with gating: "unstable" — advisory. It set the build UNSTABLE but does not block the merge. Worth a look, not a stop.

spyre-ci · 2026-09-10T07:32:51Z

GHA test runs:

[torch-spyre/x86_64] integration: DISPATCHED https://github.com/torch-spyre/torch-spyre/actions/runs/34450495476

spyre-ci · 2026-09-10T07:34:59Z

GHA test runs:

[hf-adapters/x86_64] regression: DISPATCHED https://github.com/torch-spyre/hf-adapters/actions/runs/34450674627

spyre-ci · 2026-09-10T07:39:32Z

GHA test runs:

[spyre-inference/x86_64] integration: DISPATCHED https://github.com/torch-spyre/spyre-inference/actions/runs/34451056047

Co-authored-by: GPT 5.6 <gpt5@users.noreply.github.com> Signed-off-by: Ariel Gera <ariel.gera1@ibm.com>

Signed-off-by: Ariel Gera <ariel.gera1@ibm.com>

Co-authored-by: GPT 5.6 <gpt5@users.noreply.github.com> Signed-off-by: Ariel Gera <ariel.gera1@ibm.com>

pksubbarao added the phase:1 label Sep 8, 2026

arielge force-pushed the gemma4_vlm_ple_moe branch 3 times, most recently from 9f3f922 to 004e763 Compare September 9, 2026 13:00

arielge marked this pull request as ready for review September 9, 2026 13:00

arielge requested review from BenjSz, Mohini-Shrivastava and kiszk as code owners September 9, 2026 13:00

arielge force-pushed the gemma4_vlm_ple_moe branch from 004e763 to 980a67a Compare September 9, 2026 13:16

arielge enabled auto-merge September 9, 2026 13:58

BenjSz previously approved these changes Sep 10, 2026

View reviewed changes

arielge added this pull request to the merge queue Sep 10, 2026

github-merge-queue Bot removed this pull request from the merge queue due to failed status checks Sep 10, 2026

arielge added this pull request to the merge queue Sep 10, 2026

github-merge-queue Bot removed this pull request from the merge queue due to failed status checks Sep 10, 2026