Add Granite-swa and Granitemoe-swa model support by daviswer · Pull Request #47179 · huggingface/transformers

daviswer · 2026-07-08T19:28:45Z

What does this PR do?

Adds two new models for upcoming Granite model releases, combining existing Granite/Granitemoe architectures respectively with per-layer sliding window attention and learnable per-head attention sinks. Supports shared experts, flash3+4 attention, and per-layer Rope/Nope toggling as well, following existing patterns (granitemoe, gpt-oss, llama4 respectively). SDPA and FlexAttention not currently supported.

Fixes #46645.

Code Agent Policy

I confirm that this is not a pure code agent PR.

Claude Code assisted in implementing these code changes; I have reviewed every line and run the HF formatting and correctness tests (SDPA, Flash2, and TP are skipped, as SDPA and Flash2 do not support sink tokens, and TP specifies SDPA).

Before submitting

This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
Did you read the contributor guideline and the
Pull Request checks?
Was this discussed/approved via a Github issue or the forum? Please add a link
to it if that's the case.
(See Add SWA+sink option to Granite models #46645)
Did you make sure to update the documentation with your changes according to the guidelines?
Did you write any new necessary tests?

Who can review?

@ArthurZucker @Cyrilvallez @vasqu

…le attention sinks - New model type 'granite_swa' with per-layer sliding window attention - LSE-based sink mechanism: sink_scale = sigmoid(lse - sinks) applied post-attention - Supports MuP parameters (embedding_multiplier, residual_multiplier, logits_scaling, attention_multiplier) - Per-layer attention pattern: configurable full_attention vs sliding_attention layers - Includes standalone modeling, configuration, modular source, and auto registration

- Uses flash_attn_interface._flash_attn_forward for FA3 path (returns LSE natively) - Falls back to eager attention with manual LSE computation - Adds _supports_flash_attn and _compatible_flash_implementations flags - Verified: both eager and FA3 produce correct output

- SWA layers store only last (sliding_window - 1) tokens in KV cache - Full attention layers keep complete KV history - ~98% KV memory savings on SWA layers for long sequences - Verified: identical results with and without sliding window cache

Make modular_granite_swa.py the single source of truth and regenerate the modeling/config: - Define GraniteSWAConfig in the modular file (inheriting GraniteConfig) so the converter no longer wipes the config, and rope_parameters standardizes correctly. - Apply the learnable attention sink through the standard `s_aux` attention dispatch (as in GPT-OSS) instead of the bespoke `_lse`/`_fa3_lse` path. Eager keeps the original sigmoid(logsumexp - sink) math; FlashAttention-3 and 4 now work via the shared interface (`s_aux`/`learnable_sink`). - Disable SDPA and FlexAttention (sink not expressible there); supported backends are eager, flash_attention_3, flash_attention_4. - Add tests/models/granite_swa (CausalLMModelTest framework) and the model doc + _toctree entry. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add GraniteMoeSWA, sublassing GraniteMoeShared, with SWA sliding window + sink added. - MoeSWAConfig inherits MoeSharedConfig, including default no shared experts (`shared_intermediate_size=0`). - Attention uses SWAAttention and thus supports Flash3+4 and eager. No SDPA or Flex. - MoeSWAModel handles per-layer causal masks during forward. All other layers inherit unchanged. - Add tests/models/granitemoe_swa (CausalLMModelTest framework), model doc and _toctree entry, and register model in auto mappings. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

vasqu

Looks good for the most part, added some comments what we could simplify imo and parts we should change (esp the tests I feel like we should enable more re TP/EP!)

vasqu · 2026-07-10T16:34:34Z

+    @unittest.skip("TP test mixin forces attn_implementation='sdpa', unsupported by GraniteMoeSWA (sink).")
+    def test_tp_forward(self):
+        pass
+
+    @unittest.skip("TP test mixin forces attn_implementation='sdpa', unsupported by GraniteMoeSWA (sink).")
+    def test_tp_backward(self):
+        pass
+
+    @unittest.skip("TP test mixin forces attn_implementation='sdpa', unsupported by GraniteMoeSWA (sink).")
+    def test_tp_generation(self):
+        pass
+
+    @unittest.skip("TP test mixin forces attn_implementation='sdpa', unsupported by GraniteMoeSWA (sink).")
+    def test_tp_generation_quantized(self):
+        pass
+
+    # Skipped for consistency with the prior Granite MoE models (granitemoe, granitemoeshared), which
+    # do not exercise the expert-parallel test path.
+    @unittest.skip("EP test path not exercised by the prior Granite MoE models; skipped for consistency.")
+    def test_ep_forward(self):
+        pass
+
+    @unittest.skip("EP test path not exercised by the prior Granite MoE models; skipped for consistency.")
+    def test_ep_backward(self):
+        pass


I'd rather override those tbh

I added tp plans so the TP tests no longer skip (and run correctly). I left the EP test exclusion in granitemoe_swa for now. LMK if you had something different in mind (not entirely sure what you meant by "override" here, sorry!)

I originally meant that if it's something specific to the model but it should run, it makes more sense to override the tests than to fully skip them

But ig TP has resolved itself :D for EP I wouldn't skip, even if it goes for prior models, we ought to skip as little as possible. Possibly we can unskip for the other granite moes as well

Sure, I'll try and get those running

-Fix header years -TP and integration tests -Mark flex support -Rope/Nope based on position embed passing -Top level model defs via modular subclassing Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

daviswer · 2026-07-15T04:05:46Z

Thanks @vasqu for the comments! I've made changes incorporating your feedback, up to 1) keeping the custom forward implementation for eager, 2) still skipping the EP tests, not entirely sure what you had in mind there? and 3) need to find out if tying rope/nope to layer type is future-safe on our end

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

github-actions · 2026-07-15T18:15:34Z

Thank you for your contribution 🤗!

CI Security Gate — automatic approval blocked

This PR was not automatically approved for CI because the security gate failed.

Possible reasons:

The PR touches 50 or more files — only PRs with fewer than 50 changed files are automatically approved
A changed file is outside the allowed directories (src/, tests/, docs/, utils/), has a disallowed extension (only .py, .txt, .md permitted outside tests/ and docs/), or is not .md/.yml inside docs/
A new high-severity security issue was detected in the changed Python files (Bandit check)

See the workflow run for the exact violations.

A maintainer can review and manually approve CI if a finding is a false positive.

daviswer · 2026-07-15T18:19:57Z

I updated the moe code so it runs against the EP tests (though isn't quite modular anymore). It also turns out that we want to support different rope thetas for different layers, regardless of attention type, not just a binary rope/nope decision, so the old 1/0 list has been replaced with a list of per-layer frequencies (with 0 for Nope, and defaulting to the global rope_theta param when None or unspecified). Current model correctness is unaffected. I think this addresses all of your first-round comments at this point @vasqu

vasqu

Looks very ready, my comments are mostly small but you uncovered that we missed EP on the granite moe model 😢 can we fix this along this PR or in a separate one? We really shouldnt need to override in this implementation for the moe portion

vasqu · 2026-07-16T15:43:40Z

+        Per-layer RoPE base (`theta`) frequency. `0` sets NoPE (no positional embedding) for
+        that layer. Overrides global `rope_parameters["rope_theta"]`, which is only used when
+        this list is not provided or specified (`layer_rope_theta = None`).
+


Oh wow so even a diff between different full attention layers? Just fmi

I think the intention is more for different SWA layers, but don't quote me on that

vasqu · 2026-07-16T15:56:43Z

+    # Attention shards like Granite (+ per-head `sinks` colwise to track the head-sharding); the
+    # routed experts shard tensor-parallel (packed gate/up colwise, down rowwise, `moe_tp_experts`)
+    # with the router replicated. The optional shared expert (`shared_mlp`, off by default) is left
+    # replicated -- it is small and its full output sums consistently with the all-reduced MoE output.
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.self_attn.sinks": "colwise",
+        "layers.*.block_sparse_moe.experts.gate_up_proj": "packed_colwise",
+        "layers.*.block_sparse_moe.experts.down_proj": "rowwise",
+        "layers.*.block_sparse_moe.experts": "moe_tp_experts",
+    }
+    # Expert-parallel plan: shard the routed experts across ranks (each rank owns a slice of the
+    # experts) with the router driving the dispatch. The optional shared expert is left replicated.
+    base_model_ep_plan = {
+        "layers.*.block_sparse_moe.router": "ep_router",
+        "layers.*.block_sparse_moe.experts.gate_up_proj": "grouped_gemm",
+        "layers.*.block_sparse_moe.experts.down_proj": "grouped_gemm",
+        "layers.*.block_sparse_moe.experts": "moe_tp_experts",
+    }


I just checked granitemoe and it seems to completely miss the tp/ep plans, I think we can at least inherit the ep plan then

See below - it appears the older Granite Moe models (#33207 #35894 #37658 ) predate the tp_plan/ep_plan logic and tests (#42697 #43722 ). So I'd prefer to update the moe router implementation for the new models, for correctness in the current env, but keep the older models grandfathered in, in their current form, if possible

vasqu · 2026-07-16T16:00:11Z

+        self.assertIsNotNone(output)
+
+    @unittest.skip("GraniteSWA does not support FlashAttention-2 (only FA3/FA4).")
+    def test_flash_attn_2_equivalence(self):


Uh oh this should be automatically disabled based on compatible fa; the other tests are skipped properly?

So something along

# Custom kernel which needs the mask interface to be properly usable on these models if not model_class._supports_attention_backend and not attn_implementation.startswith("flash_attention"): continue

It looks like there's two different flash 2 tests: ModelTesterMixin.test_flash_attn_2_inference_equivalence from test_modeling_common.py, and CausalLMModelTest.test_flash_attn_2_equivalence from causal_lm_tester.py. The former does automatically disable based on supported flash attentions, but the latter only skips when flash is not supported at all, so we (and gpt-oss) skip that one manually. In our case, the test does actually pass because FA2 gets auto-corrected to FA3 during model construction, but now the test isn't really testing FA2 equivalence anymore. I can keep the skip with an updated comment or just let it run, lmk what you think would be better

I'd rather skip properly as in the mixin, does was not supposed to fallback for these tests otherwise the meaning of these tests get messy

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

vasqu · 2026-07-20T15:21:44Z

Checking now!

vasqu

Small comments left but close to merging 🤗

vasqu · 2026-07-20T15:32:13Z

+        # Per-layer RoPE base theta (0 => NoPE). Default: global rope_theta
+        if self.layer_rope_theta is None:
+            self.layer_rope_theta = [self.rope_parameters["rope_theta"]] * self.num_hidden_layers


Nit: I would move this before the super call if possible

Not really feasible unfortunately - the super call sets / makes consistent the self.rope_parameters["rope_theta"] field

vasqu · 2026-07-20T15:57:53Z

+    embedding_multiplier (`float`, *optional*, defaults to 1.0):
+        embedding multiplier
+    logits_scaling (`float`, *optional*, defaults to 1.0):
+        divisor for output logits
+    residual_multiplier (`float`, *optional*, defaults to 1.0):
+        residual multiplier
+    attention_multiplier (`float`, *optional*, defaults to 1.0):
+        attention multiplier


A lot of multipliers that are actually not doing anything (based on the default values). Are they really necessary for this architecture?

Could we remove

Or at least shorten where we don't need them

If it's for modular it's fine as well. Just for my clarification

The multiplier fields are necessary because we publish models with non-default values, but the docstring entries aren't needed here and I'll remove them

vasqu · 2026-07-20T16:07:44Z

+        # Mirror CausalLMModelTest's Flash2 check; otherwise FA2 auto-corrects to FA3 and no longer tests FA2.
+        if "flash_attention_2" not in (self.model_tester.base_model_class._compatible_flash_implementations or []):
+            self.skipTest("flash_attention_2 not in _compatible_flash_implementations (FA3/FA4-only)")
+        super().test_flash_attn_2_equivalence()


Ah no I actually meant to add this to the causallm tester. This should never happen for any of these models (with the same skip reason as the base mixin)

# Some models only support a sub set of all FA implementations valid_fa_implementations = model._compatible_flash_implementations if valid_fa_implementations is not None and attn_implementation not in valid_fa_implementations: continue

ah ok, I wasn't sure if the shared tester harness was a valid target for a new-model PR

It usually isn't but sometimes we do hit those cases :D

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

github-actions · 2026-07-29T17:34:49Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, granite_swa, granitemoe_swa

vasqu · 2026-07-29T17:35:33Z

run-slow: granite_swa, granitemoe_swa

vasqu · 2026-07-29T17:36:46Z

@daviswer I updated the values here 5474d5b - not 100% sure if the repetition is expected on the moe swa variation.

github-actions · 2026-07-29T17:37:10Z

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/granite_swa", "models/granitemoe_swa"]
quantizations: []

github-actions · 2026-07-29T17:38:55Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	97e985e7	workflow commit (merge commit)
PR	8b242a63	branch commit (from PR)
main	2ef79f87	base commit (on `main`)

⚠️ No test being reported (jobs are skipped or cancelled)!

vasqu · 2026-07-29T17:40:08Z

run-slow: granite_swa, granitemoe_swa

github-actions · 2026-07-29T17:41:39Z

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/granite_swa", "models/granitemoe_swa"]
quantizations: []

github-actions · 2026-07-29T17:53:28Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	6eb2dc42	workflow commit (merge commit)
PR	d4cf007c	branch commit (from PR)
main	2ef79f87	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

HuggingFaceDocBuilderDev · 2026-07-29T17:54:00Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

daviswer · 2026-07-29T20:24:20Z

Thanks @vasqu , the repetition is a bit surprising but it's not a correctness issue, the reference model is just somewhat experimental.

FWIW the old output should still be correct for both cuda 8 and 9, but if adding that back in means rerunning the slow tests again then feel free to skip

vasqu · 2026-07-29T21:03:46Z

I think it's just a subtle difference between A100 and A10 GPUs, it can happen, I'd merge then 🫡

daviswer · 2026-07-29T22:28:54Z

Great, thanks so much for all your help with this @vasqu !

* Add GraniteSWA model: Granite with Sliding Window Attention + learnable attention sinks - New model type 'granite_swa' with per-layer sliding window attention - LSE-based sink mechanism: sink_scale = sigmoid(lse - sinks) applied post-attention - Supports MuP parameters (embedding_multiplier, residual_multiplier, logits_scaling, attention_multiplier) - Per-layer attention pattern: configurable full_attention vs sliding_attention layers - Includes standalone modeling, configuration, modular source, and auto registration * Add FA3 attention dispatch with native LSE for sink scaling - Uses flash_attn_interface._flash_attn_forward for FA3 path (returns LSE natively) - Falls back to eager attention with manual LSE computation - Adds _supports_flash_attn and _compatible_flash_implementations flags - Verified: both eager and FA3 produce correct output * Enable DynamicSlidingWindowCache for memory-efficient KV cache - SWA layers store only last (sliding_window - 1) tokens in KV cache - Full attention layers keep complete KV history - ~98% KV memory savings on SWA layers for long sequences - Verified: identical results with and without sliding window cache * Finalize GraniteSWA: modular impl, FA4 support, formatting, tests, docs Make modular_granite_swa.py the single source of truth and regenerate the modeling/config: - Define GraniteSWAConfig in the modular file (inheriting GraniteConfig) so the converter no longer wipes the config, and rope_parameters standardizes correctly. - Apply the learnable attention sink through the standard `s_aux` attention dispatch (as in GPT-OSS) instead of the bespoke `_lse`/`_fa3_lse` path. Eager keeps the original sigmoid(logsumexp - sink) math; FlashAttention-3 and 4 now work via the shared interface (`s_aux`/`learnable_sink`). - Disable SDPA and FlexAttention (sink not expressible there); supported backends are eager, flash_attention_3, flash_attention_4. - Add tests/models/granite_swa (CausalLMModelTest framework) and the model doc + _toctree entry. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Add GraniteMoeSWA (MoeShared+SWA) Add GraniteMoeSWA, sublassing GraniteMoeShared, with SWA sliding window + sink added. - MoeSWAConfig inherits MoeSharedConfig, including default no shared experts (`shared_intermediate_size=0`). - Attention uses SWAAttention and thus supports Flash3+4 and eager. No SDPA or Flex. - MoeSWAModel handles per-layer causal masks during forward. All other layers inherit unchanged. - Add tests/models/granitemoe_swa (CausalLMModelTest framework), model doc and _toctree entry, and register model in auto mappings. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Update model refs, add per-layer rope/nope support a la Llama4 Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Small formatting/date fixes Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Test comments Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * automodel import style fix * Regenerate modeling_ files post-resync to main Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Fix upload date in model cards Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Additional TP / historic test exclusions post-sync to main Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Docs cleanup, one final round on submit date Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Fix grante_swa init and imports Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Addressing review comments: -Fix header years -TP and integration tests -Mark flex support -Rope/Nope based on position embed passing -Top level model defs via modular subclassing Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Replace binary rope/nope support with dynamic rope freq support Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Reorder router outputs to support native EP (and tests) Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Clean flex flags, flash correction tests, rope constructors Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Correct Flash2 test skip check Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Clean docstrings/comments, fix flash2 causal lm test Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Modelcard dates Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Fix rope typing and slow test shape checking Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Update integration test expectations by device Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * Regenerate modeling files to match inherited modular upstream changes Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> * adjust to ci * fixup repo * try * kernels sync --------- Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com> Co-authored-by: Bharat-Runwal <bharatrunwal@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: vasqu <antonprogamer@gmail.com> Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>

Bharat-Runwal and others added 16 commits May 28, 2026 09:45

Update model refs, add per-layer rope/nope support a la Llama4

56a2274

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Small formatting/date fixes

c40cf93

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Test comments

ddada04

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Merge branch 'main' into granite-swa

76d88f7

automodel import style fix

da94124

Fix automodel import order after pull

f77b683

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Regenerate modeling_ files post-resync to main

e9baf4d

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Fix upload date in model cards

047cd7e

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Additional TP / historic test exclusions post-sync to main

6a6f053

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Docs cleanup, one final round on submit date

4e99d5a

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Fix grante_swa init and imports

e0d10d4

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

gabe-l-hart mentioned this pull request Jul 9, 2026

model: GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM ggml-org/llama.cpp#25505

Merged

vasqu reviewed Jul 10, 2026

View reviewed changes

Addressing review comments:

e04cd0b

-Fix header years -TP and integration tests -Mark flex support -Rope/Nope based on position embed passing -Top level model defs via modular subclassing Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

daviswer added 2 commits July 15, 2026 16:36

Replace binary rope/nope support with dynamic rope freq support

91e6663

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Reorder router outputs to support native EP (and tests)

c984f0b

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

vasqu reviewed Jul 16, 2026

View reviewed changes

daviswer added 2 commits July 17, 2026 16:43

Clean flex flags, flash correction tests, rope constructors

86b5716

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

Correct Flash2 test skip check

5f04c02

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

vasqu approved these changes Jul 20, 2026

View reviewed changes

Clean docstrings/comments, fix flash2 causal lm test

1b27b6b

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

daviswer and others added 2 commits July 28, 2026 21:30

Regenerate modeling files to match inherited modular upstream changes

a866b4a

Signed-off-by: Davis Wertheimer <davis.wertheimer@ibm.com>

adjust to ci

5474d5b

vasqu and others added 2 commits July 29, 2026 19:35

fixup repo

2cee0ae

Merge branch 'main' into granite-swa

8b242a6

try

d4cf007

kernels sync

c3ba1d4

vasqu enabled auto-merge July 29, 2026 21:03

vasqu added this pull request to the merge queue Jul 29, 2026

github-merge-queue Bot removed this pull request from the merge queue due to failed status checks Jul 29, 2026

vasqu added this pull request to the merge queue Jul 29, 2026

Merged via the queue into huggingface:main with commit d24918c Jul 29, 2026
114 checks passed

github-actions Bot mentioned this pull request Aug 17, 2026

⚙️ AI Infrastructure Digest 2026-08-18 Neare-Design/agents-radar#136

Open

github-actions Bot mentioned this pull request Aug 26, 2026

[Genesis Task] new capability — Autonomously add one bounded executable Genesis capability named learned_15ea4d08866af6f5. Use the learned idea: Add one new bounded Genesis capabilit Maxhm007/Genesis-AI-Network#412

Open

sergereview Bot mentioned this pull request Sep 4, 2026

[serge] Fix 4 integration tests for model blt failing with import_or_config (other (4)) #48534

Closed

sergereview Bot mentioned this pull request Sep 27, 2026

[serge] Fix 2 integration tests for model jais2 failing with output_mismatch (tensor values differ (2)) #49143

Open

Conversation

daviswer commented Jul 8, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Code Agent Policy

Before submitting

Who can review?

Uh oh!

vasqu left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

daviswer commented Jul 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Jul 15, 2026

CI Security Gate — automatic approval blocked

Uh oh!

daviswer commented Jul 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vasqu left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

daviswer Jul 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

vasqu commented Jul 20, 2026

Uh oh!

vasqu left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

daviswer commented Jul 8, 2026 •

edited by github-actions Bot

Loading

daviswer commented Jul 15, 2026 •

edited

Loading

daviswer commented Jul 15, 2026 •

edited

Loading

daviswer Jul 17, 2026 •

edited

Loading

daviswer commented Jul 29, 2026 •

edited

Loading