[None][feat] Support post-norm and per-aux fc_norm for Eagle3 draft models by Dogacel · Pull Request #14988 · NVIDIA/TensorRT-LLM

Dogacel · 2026-06-05T04:44:12Z

Description

Adds support for EAGLE-3.1 architecture & draft checkpoints. Related sources:

Paper describing the architectural changes + rational: https://arxiv.org/abs/2605.09992
vLLM blog post on EAGLE 3.1: https://vllm.ai/blog/2026-05-26-eagle-3-1
Newly supported models: https://huggingface.co/Dogacel/specdrift-gpt-oss-120b-eagle3, https://huggingface.co/Dogacel/specdrift-gpt-oss-20b-eagle3, https://huggingface.co/lightseekorg/kimi-k2.6-eagle3.1-mla (not tested)

Core changes:

norm_output (post-norm): return the post-final-norm hidden state as the auxiliary feature fed to the next draft step.
fc_norm (per-aux norm): apply a separate RMSNorm to each captured hidden state before the fc projection. Unlike the existing single-norm norm_before_fc (one norm over the full concatenated vector), this normalizes each of the num_capture_layers features independently so they contribute equally regardless of raw scale.

The change is additive and behavior-preserving for existing drafters.

Results

gpt-oss-120b, 2×H100, max_draft_len: 7, greedy. Values are tokens / forward pass (accepted_len = value − 1; 1.00 = no speculation, 8.00 = theoretical max).

prompt	baseline (NVIDIA) low / high	EAGLE3.1 (ours) low / high
trivia_1line	2.40 / 2.48	5.00 / 3.93
factual_para	1.59 / 1.81	1.98 / 2.53
code_gen	2.29 / 1.99	3.35 / 2.71
math_reason	2.44 / 2.07	4.03 / 3.85
logic_puzzle	2.00 / 2.21	3.36 / 4.34
mean tok/forward	2.13	3.51
mean accepted_len	~1.13	~2.51

Models:

NVIDIA: https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3
Specdrift (Ours EAGLE 3.1): https://huggingface.co/Dogacel/specdrift-gpt-oss-120b-eagle3

Testing Strategy

Both models tested using openai/gpt-oss-120b on 2×H100 with --tp_size 2, differing only in the drafter + config.

regular.sh (baseline — stock NVIDIA 3-layer drafter, used to confirm no regression):

trtllm-serve openai/gpt-oss-120b \
    --host 0.0.0.0 --port 8888 \
    --backend pytorch \
    --max_batch_size 32 --max_num_tokens 8192 --max_seq_len 8192 \
    --tp_size 2 \
    --extra_llm_api_options /workdir/extra-llm-api-config.yml

extra-llm-api-config.yml:

enable_attention_dp: false
disable_overlap_scheduler: true
enable_autotuner: false
return_perf_metrics: true
perf_metrics_max_requests: 4096
cuda_graph_config:
  max_batch_size: 1
speculative_config:
  decoding_type: Eagle
  max_draft_len: 7
  speculative_model_dir: nvidia/gpt-oss-120b-Eagle3-v3
kv_cache_config:
  enable_block_reuse: false

new.sh (new model architecture supported using this PR):

trtllm-serve openai/gpt-oss-120b \
    --host 0.0.0.0 --port 8888 \
    --backend pytorch \
    --max_batch_size 32 --max_num_tokens 8192 --max_seq_len 8192 \
    --tp_size 2 \
    --extra_llm_api_options /workdir/new-eagle3-config.yml

new-eagle3-config.yml (the 5 capture layers come from the drafter's eagle_aux_hidden_state_layer_ids; wiring them here makes num_capture_layers == 5, which drives the fc input dim, the fc_norm count, and the target-model capture points):

enable_attention_dp: false
disable_overlap_scheduler: true
enable_autotuner: false
return_perf_metrics: true
perf_metrics_max_requests: 4096
cuda_graph_config:
  max_batch_size: 1
speculative_config:
  decoding_type: Eagle
  max_draft_len: 7
  speculative_model_dir: dogacel/specdrift-gpt-oss-120b-eagle3
  eagle3_layers_to_capture: [1, 9, 17, 25, 33]
kv_cache_config:
  enable_block_reuse: false

Speculative decoding acceptance length testing script used (Mostly AI generated):

#!/usr/bin/env python3
"""Sweep prompt type x reasoning_effort and report Eagle3 acceptance (tokens/forward)."""
import json
import sys
import time

import requests

URL = "http://localhost:8888"
MODEL = "openai/gpt-oss-120b"
MAX_TOKENS = 512

PROMPTS = {
    "trivia_1line":  "What is the capital of France?",
    "factual_para":  "Write 150 words about the history of the bicycle.",
    "code_gen":      "Write a Python function `binary_search(arr, target)` with a "
                     "docstring, type hints, and handle the empty-list case.",
    "math_reason":   "Compute the determinant of [[2,1,3],[0,4,1],[5,2,1]] showing "
                     "every step of cofactor expansion.",
    "logic_puzzle":  "A farmer has 17 sheep. All but 9 run away. How many remain? "
                     "Reason carefully and double-check before answering.",
}
EFFORTS = ["low", "high"]


def drain():
    try:
        requests.get(f"{URL}/perf_metrics", timeout=5)
    except requests.RequestException:
        pass


def probe(prompt, effort):
    drain()
    body = {
        "model": MODEL, "temperature": 0, "max_tokens": MAX_TOKENS,
        "reasoning_effort": effort,
        "messages": [{"role": "user", "content": prompt}],
    }
    t0 = time.perf_counter()
    r = requests.post(f"{URL}/v1/chat/completions", json=body, timeout=300)
    dt = time.perf_counter() - t0
    r.raise_for_status()
    d = r.json()
    msg = d["choices"][0]["message"]
    ctok = d["usage"]["completion_tokens"]
    reasoning_tok = len((msg.get("reasoning_content") or "").split())
    time.sleep(0.3)
    pm = requests.get(f"{URL}/perf_metrics", timeout=5).json()
    steps = 0
    for e in pm:
        sm = (e.get("time_breakdown_metrics") or {}).get("step_metrics")
        if sm:
            steps += len(sm)
        else:
            p = e.get("perf_metrics") or {}
            if p.get("last_iter") is not None:
                steps += p["last_iter"] - p["first_iter"]
    tpf = ctok / steps if steps else float("nan")
    return tpf, ctok, steps, reasoning_tok, ctok / dt if dt else 0


def main():
    try:
        if requests.get(f"{URL}/health", timeout=3).status_code != 200:
            raise RuntimeError
    except Exception:
        print(f"[!] no healthy server at {URL}")
        sys.exit(1)

    print(f"{'prompt':<14}{'effort':<7}{'tok/fwd':>8}{'acc_len':>8}"
          f"{'~reason_wd':>11}{'ctok':>7}{'steps':>7}{'tok/s':>8}")
    print("-" * 70)
    best = (None, -1)
    rows = []
    for name, prompt in PROMPTS.items():
        for effort in EFFORTS:
            try:
                tpf, ctok, steps, rtok, tps = probe(prompt, effort)
            except requests.RequestException as e:
                print(f"{name:<14}{effort:<7}  ERROR {e}")
                continue
            rows.append((name, effort, tpf))
            print(f"{name:<14}{effort:<7}{tpf:>8.2f}{tpf-1:>8.2f}"
                  f"{rtok:>11}{ctok:>7}{steps:>7}{tps:>8.1f}")
            if tpf > best[1]:
                best = (f"{name}/{effort}", tpf)

    print("-" * 70)
    if rows:
        avg = sum(r[2] for r in rows) / len(rows)
        print(f"mean tok/forward across configs: {avg:.2f}")
        print(f"best: {best[0]} @ {best[1]:.2f} tok/forward "
              f"(accepted_len {best[1]-1:.2f})")


if __name__ == "__main__":
    main()

PR Checklist

Please review the following before submitting your PR:

PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.

Summary by CodeRabbit

New Features
- Added optional per-capture-layer normalization support for Eagle3 draft models, configurable through settings.
Improvements
- Enhanced hidden state processing during forward pass with conditional normalization application based on model configuration.

…odels Enable SGLang-style Eagle3 draft checkpoints in the PyTorch backend: - norm_output (post-norm): return the post-final-norm hidden state as the auxiliary feature fed to the next draft step, in addition to the existing eagle_config return_hidden_post_norm flag. - fc_norm: per-aux-layer RMSNorm applied to each captured hidden state before the fc projection. Unlike the existing single-norm norm_before_fc, this normalizes each of the num_capture_layers features independently. Combined with the configurable num_capture_layers (eagle3_layers_to_capture), this allows running drafters with 5 aux capture layers such as dogacel/specdrift-gpt-oss-120b-eagle3. Signed-off-by: Doğaç Eldenk <dogacel@gmail.com>

coderabbitai · 2026-06-05T05:03:30Z

📝 Walkthrough

Walkthrough

Eagle3 draft model implementation adds optional per-capture-layer normalization through fc_norm. Configuration determines when normalization applies; initialization creates RMSNorm layers per capture layer when enabled; the forward pass chunks hidden states and applies corresponding normalization before projection.

Changes

Eagle3 per-capture-layer normalization

Layer / File(s)	Summary
fc_norm configuration and module initialization `tensorrt_llm/_torch/models/modeling_speculative.py`	`Eagle3DraftModel` computes `_return_hidden_post_norm` from `eagle_config` or `config.norm_output`, and conditionally creates `fc_norm` as a `ModuleList` of `RMSNorm` instances—one per `spec_config.num_capture_layers`—when `config.fc_norm` is enabled.
fc_norm application in forward path `tensorrt_llm/_torch/models/modeling_speculative.py`	`Eagle3ForCausalLM.apply_eagle3_fc()` checks for `model.fc_norm` presence; when set and hidden-state projection is required, it chunks `hidden_states` along the last dimension per `fc_norm` length, applies the corresponding `RMSNorm` to each chunk, and concatenates normalized chunks before projection.

🎯 2 (Simple) | ⏱️ ~12 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is fully related to the main change in the changeset, clearly summarizing the two core features (post-norm and per-aux fc_norm) added to support Eagle3 draft models.
Description check	✅ Passed	The description is comprehensive and well-structured, covering the motivation (with references), core changes, performance results, testing strategy with detailed examples, and addressing the PR checklist items.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

Warning

Review ran into problems

🔥 Problems

Stopped waiting for pipeline failures after 30000ms. One of your pipelines takes longer than our 30000ms fetch window to run, so review may not consider pipeline-failure results for inline comments if any failures occurred after the fetch window. Increase the timeout if you want to wait longer or run a @coderabbit review after the pipeline has finished.

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coderabbitai

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 `@tensorrt_llm/_torch/models/modeling_speculative.py`:
- Around line 606-613: The zip used when applying per-chunk normalization over
self.model.fc_norm and chunks can silently truncate if lengths diverge; update
the comprehension in the fc_norm branch (where hidden_states is split with
hidden_states.chunk and norms applied via for norm, chunk in
zip(self.model.fc_norm, chunks)) to call zip(self.model.fc_norm, chunks,
strict=True) so mismatched lengths raise an error, ensuring strict pairing
between norm layers and chunks.

🪄 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: CHILL

Plan: Enterprise

Run ID: b4434610-3b52-4e63-83bc-0523f95a676f

📥 Commits

Reviewing files that changed from the base of the PR and between 81e86a5 and 0ac8947.

📒 Files selected for processing (1)

tensorrt_llm/_torch/models/modeling_speculative.py

benchislett · 2026-06-05T15:24:26Z

@mikeiovine Is there a bug in the support for https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3? The reported mean acceptance length seems extremely low
@Dogacel Could you please run evals on an actual dataset, MTBench and/or SPEED-Bench, so that we can compare the results against the model cards?

mikeiovine · 2026-06-05T16:22:36Z

@mikeiovine Is there a bug in the support for https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3? The reported mean acceptance length seems extremely low

Will check. For now I think this change is fine to land as code changes are straightforward.

mikeiovine · 2026-06-05T16:34:34Z

We should run a round of SPEED-bench with both the new drafters and https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3.

https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3 has expected ALs in different categories. We can compare to those numbers to figure out if there are bugs (I don't expect any issues though).

Dogacel · 2026-06-05T17:55:27Z

@mikeiovine Is there a bug in the support for https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3? The reported mean acceptance length seems extremely low

@Dogacel Could you please run evals on an actual dataset, MTBench and/or SPEED-Bench, so that we can compare the results against the model cards?

I've re-run the benchmark on MT-Bench, scoring 2.8 AL matching other implementations & expectation.

mikeiovine · 2026-06-05T20:05:30Z

2.8 AL on https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3, right? What does the new drafter get?

Dogacel · 2026-06-05T20:38:45Z

2.8 AL on https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-v3, right? What does the new drafter get?

Oh sorry for not clarifying. I've only tested my drafter (specdrift). I run the benchmark to validate the implementation is correct and we get the expected acceptance length.

I think testing NVIDIA's is out of scope for this PR. I've only run it as a dry run to validate things are not entirely broken. I previously tested that in vLLM and results were similar to our model in standard benchmarks. The difference is visible in OOD cases.

hongyanz · 2026-06-05T21:30:20Z

Hi @mikeiovine , let's focus on merging this PR only. As the codes are all open sourced, please feel free to run any tests that you are interested in at your side, as different machines may vary the performance too. Thank you.

mikeiovine · 2026-06-11T00:42:23Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-11T00:49:21Z

PR_Github #53410 [ run ] triggered by Bot. Commit: d7b4ac1 Link to invocation

tensorrt-cicd · 2026-06-11T07:54:04Z

PR_Github #53410 [ run ] completed with state SUCCESS. Commit: d7b4ac1
/LLM/main/L0_MergeRequest_PR pipeline #42583 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-11T21:37:00Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-11T21:42:45Z

PR_Github #53688 [ run ] triggered by Bot. Commit: 3f390e6 Link to invocation

tensorrt-cicd · 2026-06-12T01:45:02Z

PR_Github #53688 [ run ] completed with state SUCCESS. Commit: 3f390e6
/LLM/main/L0_MergeRequest_PR pipeline #42825 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-12T15:29:25Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-12T15:35:44Z

PR_Github #53906 [ run ] triggered by Bot. Commit: 6d774b9 Link to invocation

mikeiovine · 2026-06-12T19:03:43Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-12T19:11:03Z

PR_Github #53941 [ run ] triggered by Bot. Commit: af33ab9 Link to invocation

tensorrt-cicd · 2026-06-12T19:13:35Z

PR_Github #53906 [ run ] completed with state ABORTED. Commit: 6d774b9

Link to invocation

tensorrt-cicd · 2026-06-12T23:08:46Z

PR_Github #53941 [ run ] completed with state FAILURE. Commit: af33ab9
/LLM/main/L0_MergeRequest_PR pipeline #43033 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-13T17:49:08Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-13T17:54:19Z

PR_Github #54064 [ run ] triggered by Bot. Commit: d61c269 Link to invocation

tensorrt-cicd · 2026-06-13T20:00:01Z

PR_Github #54064 [ run ] completed with state SUCCESS. Commit: d61c269
/LLM/main/L0_MergeRequest_PR pipeline #43148 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-16T00:08:15Z

/bot run --disable-fail-fast

mikeiovine · 2026-06-16T00:09:30Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-16T00:15:13Z

PR_Github #54383 [ run ] triggered by Bot. Commit: 7fa16ca Link to invocation

tensorrt-cicd · 2026-06-16T00:15:45Z

PR_Github #54384 [ run ] triggered by Bot. Commit: 7fa16ca Link to invocation

tensorrt-cicd · 2026-06-16T00:15:52Z

PR_Github #54383 [ run ] completed with state ABORTED. Commit: 7fa16ca

Link to invocation

tensorrt-cicd · 2026-06-16T04:10:58Z

PR_Github #54384 [ run ] completed with state FAILURE. Commit: 7fa16ca
/LLM/main/L0_MergeRequest_PR pipeline #43454 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-16T16:29:05Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-16T16:38:24Z

PR_Github #54624 [ run ] triggered by Bot. Commit: 0f9ef99 Link to invocation

tensorrt-cicd · 2026-06-16T20:50:01Z

PR_Github #54624 [ run ] completed with state FAILURE. Commit: 0f9ef99
/LLM/main/L0_MergeRequest_PR pipeline #43673 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-22T14:54:32Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-22T15:01:02Z

PR_Github #55045 [ run ] triggered by Bot. Commit: 03f46b9 Link to invocation

tensorrt-cicd · 2026-06-22T19:14:17Z

PR_Github #55045 [ run ] completed with state SUCCESS. Commit: 03f46b9
/LLM/main/L0_MergeRequest_PR pipeline #44036 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-23T15:11:48Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-23T15:17:35Z

PR_Github #55254 [ run ] triggered by Bot. Commit: 36da7de Link to invocation

tensorrt-cicd · 2026-06-23T19:50:44Z

PR_Github #55254 [ run ] completed with state SUCCESS. Commit: 36da7de
/LLM/main/L0_MergeRequest_PR pipeline #44213 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

mikeiovine · 2026-06-23T20:14:45Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-23T20:20:15Z

PR_Github #55326 [ run ] triggered by Bot. Commit: 9b0dd64 Link to invocation

github-actions Bot assigned Dogacel Jun 5, 2026

Dogacel marked this pull request as ready for review June 5, 2026 04:55

Dogacel requested a review from a team as a code owner June 5, 2026 04:55

Dogacel requested a review from yechank-nvidia June 5, 2026 04:55

coderabbitai Bot reviewed Jun 5, 2026

View reviewed changes

Comment thread tensorrt_llm/_torch/models/modeling_speculative.py

mikeiovine approved these changes Jun 5, 2026

View reviewed changes

Merge branch 'main' into eagle3_1-postnorm

d7b4ac1

laikhtewari enabled auto-merge (squash) June 10, 2026 23:03

Merge branch 'main' into eagle3_1-postnorm

3f390e6

Merge branch 'main' into eagle3_1-postnorm

6d774b9

Merge branch 'main' into eagle3_1-postnorm

af33ab9

Merge branch 'main' into eagle3_1-postnorm

d61c269

yechank-nvidia approved these changes Jun 15, 2026

View reviewed changes

Merge branch 'main' into eagle3_1-postnorm

7fa16ca

Merge branch 'main' into eagle3_1-postnorm

0f9ef99

Merge branch 'main' into eagle3_1-postnorm

03f46b9

Merge branch 'main' into eagle3_1-postnorm

36da7de

Merge branch 'main' into eagle3_1-postnorm

9b0dd64

Conversation

Dogacel commented Jun 5, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Results

Testing Strategy

PR Checklist

Summary by CodeRabbit

Uh oh!

coderabbitai Bot commented Jun 5, 2026

Walkthrough

Changes

❌ Failed checks (1 warning)

Review ran into problems

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

benchislett commented Jun 5, 2026

Uh oh!

mikeiovine commented Jun 5, 2026

Uh oh!

mikeiovine commented Jun 5, 2026

Uh oh!

Dogacel commented Jun 5, 2026

Uh oh!

mikeiovine commented Jun 5, 2026

Uh oh!

Dogacel commented Jun 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hongyanz commented Jun 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mikeiovine commented Jun 11, 2026

Uh oh!

tensorrt-cicd commented Jun 11, 2026

Uh oh!

tensorrt-cicd commented Jun 11, 2026

Uh oh!

mikeiovine commented Jun 11, 2026

Uh oh!

tensorrt-cicd commented Jun 11, 2026

Uh oh!

tensorrt-cicd commented Jun 12, 2026

Uh oh!

mikeiovine commented Jun 12, 2026

Uh oh!

tensorrt-cicd commented Jun 12, 2026

Uh oh!

mikeiovine commented Jun 12, 2026

Uh oh!

tensorrt-cicd commented Jun 12, 2026

Uh oh!

tensorrt-cicd commented Jun 12, 2026

Uh oh!

tensorrt-cicd commented Jun 12, 2026

Uh oh!

mikeiovine commented Jun 13, 2026

Uh oh!

tensorrt-cicd commented Jun 13, 2026

Uh oh!

tensorrt-cicd commented Jun 13, 2026

Uh oh!

mikeiovine commented Jun 16, 2026

Uh oh!

mikeiovine commented Jun 16, 2026

Uh oh!

tensorrt-cicd commented Jun 16, 2026

Uh oh!

tensorrt-cicd commented Jun 16, 2026

Uh oh!

tensorrt-cicd commented Jun 16, 2026

Uh oh!

tensorrt-cicd commented Jun 16, 2026

Uh oh!

mikeiovine commented Jun 16, 2026

Uh oh!

tensorrt-cicd commented Jun 16, 2026

Dogacel commented Jun 5, 2026 •

edited by coderabbitai Bot

Loading

Dogacel commented Jun 5, 2026 •

edited

Loading

hongyanz commented Jun 5, 2026 •

edited

Loading