From 041c758436976a28b13207de622d68d2877ea0e8 Mon Sep 17 00:00:00 2001
From: noeyy-mino <174223378+noeyy-mino@users.noreply.github.com>
Date: Thu, 16 Jul 2026 07:01:47 +0000
Subject: [PATCH 1/2] launcher: bump TRT-LLM to 1.3.0rc20, pin vLLM to v0.22.0,
 fix max_seq_len for Qwen3.5-4B

Signed-off-by: noeyy-mino <174223378+noeyy-mino@users.noreply.github.com>
---
 .../Qwen/Qwen3-30B-A3B/hf_streaming_dflash_multi_node.yaml  | 2 +-
 .../Qwen/Qwen3-30B-A3B/hf_streaming_eagle3_multi_node.yaml  | 2 +-
 tools/launcher/examples/Qwen/Qwen3-8B/hf_eagle3_dryrun.yaml | 2 +-
 tools/launcher/examples/Qwen/Qwen3-8B/hf_online_eagle3.yaml | 4 ++--
 .../examples/Qwen/Qwen3-8B/hf_streaming_dflash.yaml         | 4 ++--
 .../Qwen/Qwen3-8B/hf_streaming_dflash_multi_node.yaml       | 4 ++--
 .../examples/Qwen/Qwen3-8B/hf_streaming_eagle3.yaml         | 2 +-
 .../Qwen/Qwen3-8B/hf_streaming_eagle3_multi_node.yaml       | 2 +-
 tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench.yaml  | 2 +-
 .../examples/Qwen/Qwen3.5-4B/specdec_bench_mtp_vllm.yaml    | 2 +-
 .../examples/moonshotai/Kimi-K2.5/hf_dflash_dryrun.yaml     | 2 +-
 .../examples/moonshotai/Kimi-K2.5/hf_eagle3_dryrun.yaml     | 2 +-
 .../examples/moonshotai/Kimi-K2.5/hf_offline_dflash.yaml    | 2 +-
 .../examples/moonshotai/Kimi-K2.5/hf_streaming_dflash.yaml  | 4 ++--
 .../Kimi-K2.5/hf_streaming_dflash_multi_node.yaml           | 4 ++--
 .../examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3.yaml  | 6 +++---
 .../Kimi-K2.5/hf_streaming_eagle3_multi_node.yaml           | 6 +++---
 .../examples/moonshotai/Kimi-K2.5/specdec_bench.yaml        | 4 ++--
 .../openai/gpt-oss-20b/hf_streaming_dflash_multi_node.yaml  | 2 +-
 .../openai/gpt-oss-20b/hf_streaming_eagle3_multi_node.yaml  | 2 +-
 tools/launcher/slurm_config.py                              | 2 +-
 21 files changed, 31 insertions(+), 31 deletions(-)

diff --git a/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_dflash_multi_node.yaml b/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_dflash_multi_node.yaml
index daefe6223d9..d5583b7cccd 100644
--- a/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_dflash_multi_node.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_dflash_multi_node.yaml
@@ -48,7 +48,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming DFlash training — 2 serve replicas (TP=2) + 2 trainer nodes (2 GPU each).
   # DFlash extracts 5 target layers (build_target_layer_ids(48,5)=[1,12,23,34,45], the
diff --git a/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_eagle3_multi_node.yaml b/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_eagle3_multi_node.yaml
index 8a7a7f2ac20..52c9fc6b16e 100644
--- a/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_eagle3_multi_node.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-30B-A3B/hf_streaming_eagle3_multi_node.yaml
@@ -44,7 +44,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming EAGLE3 training — 2 serve replicas (TP=2) + 2 trainer nodes (2 GPU each).
   # Capture ids: default_eagle_aux_layer_ids(48)=[1,23,44] +1, plus final layer 48.
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_eagle3_dryrun.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_eagle3_dryrun.yaml
index 7bdba369467..7fc590cba93 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_eagle3_dryrun.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_eagle3_dryrun.yaml
@@ -53,4 +53,4 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_online_eagle3.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_online_eagle3.yaml
index 89a578ad57c..15e6d3a5f61 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_online_eagle3.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_online_eagle3.yaml
@@ -31,7 +31,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   task_1:
     script: common/eagle3/train_eagle.sh
@@ -49,7 +49,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 8
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   task_2:
     script: common/specdec_bench/quick_check.sh
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash.yaml
index 518878f5c6d..def98cf7f5f 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash.yaml
@@ -36,7 +36,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming DFlash training — node 0 vllm serve, node 1 trainer.
   # DFlash extracts 5 target layers (build_target_layer_ids(36,5)=[1,9,17,25,33], the
@@ -90,7 +90,7 @@ pipeline:
       - MIN_ACCEPTANCE_LENGTH: "1.2"
     slurm_config:
       _factory_: "slurm_factory"
-      container: "vllm/vllm-openai:nightly"
+      container: vllm/vllm-openai:nightly
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash_multi_node.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash_multi_node.yaml
index fad36cb7d98..ee8894b4b57 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash_multi_node.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_dflash_multi_node.yaml
@@ -34,7 +34,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming DFlash training — 2 serve replicas (TP=2) + 2 trainer nodes (2 GPU each).
   # DFlash extracts 5 target layers (build_target_layer_ids(36,5)=[1,9,17,25,33], the
@@ -93,7 +93,7 @@ pipeline:
       - MIN_ACCEPTANCE_LENGTH: "1.2"
     slurm_config:
       _factory_: "slurm_factory"
-      container: "vllm/vllm-openai:nightly"
+      container: vllm/vllm-openai:nightly
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3.yaml
index 59656709d93..b24468fe339 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3.yaml
@@ -30,7 +30,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # capture ids = default_eagle_aux_layer_ids(36)=[1,17,32] shifted +1, plus final
   # layer 36 -> [2,18,33,36].
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3_multi_node.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3_multi_node.yaml
index ccf8a9c8f8c..1e0dab95a35 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3_multi_node.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_streaming_eagle3_multi_node.yaml
@@ -31,7 +31,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming EAGLE3 training — 2 serve replicas (TP=2) + 2 trainer nodes (2 GPU each).
   # Capture ids: default_eagle_aux_layer_ids(36)=[1,17,32] +1, plus final layer 36.
diff --git a/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench.yaml b/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench.yaml
index d6872339f8c..318e861121e 100644
--- a/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench.yaml
@@ -76,7 +76,7 @@ pipeline:
       - --concurrency 8
       - --num_requests 80
       - --output_length 4096
-      - --max_seq_len 40960
+      - --max_seq_len 65536
       - --aa_timing
       - --show_progress
       - --save_dir /scratchspace/qwen35_4b_none_vllm/throughput_32k
diff --git a/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench_mtp_vllm.yaml b/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench_mtp_vllm.yaml
index 93b36fe74d0..7c94c8dc33f 100644
--- a/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench_mtp_vllm.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench_mtp_vllm.yaml
@@ -51,7 +51,7 @@ pipeline:
       - --ep_size 1
       - --concurrency 8
       - --num_requests 80
-      - --runtime_params common/specdec_bench/runtime_params_throughput_32k.yaml
+      - --max_seq_len 65536
       - --output_length 4096
       - --aa_timing
       - --show_progress
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_dflash_dryrun.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_dflash_dryrun.yaml
index 5cb467b3f6a..f688e6c7d6e 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_dflash_dryrun.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_dflash_dryrun.yaml
@@ -36,4 +36,4 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_eagle3_dryrun.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_eagle3_dryrun.yaml
index 9f87e404f19..b59edd01c66 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_eagle3_dryrun.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_eagle3_dryrun.yaml
@@ -57,4 +57,4 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_offline_dflash.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_offline_dflash.yaml
index 709d2cba900..17de758f36d 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_offline_dflash.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_offline_dflash.yaml
@@ -39,4 +39,4 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 8
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash.yaml
index 6dbce4eb757..32c1efc237a 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash.yaml
@@ -48,7 +48,7 @@ pipeline:
       ntasks_per_node: 1
       # The cluster QOS requires whole-node GPU allocation though make_dataset is CPU-only.
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   task_1:
     script: common/eagle3/train_eagle_streaming.sh
@@ -96,4 +96,4 @@ pipeline:
       segment: 2
       ntasks_per_node: 1
       gpus_per_node: 4
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash_multi_node.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash_multi_node.yaml
index b24dd04c458..64d13b8a503 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash_multi_node.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_dflash_multi_node.yaml
@@ -45,7 +45,7 @@ pipeline:
       ntasks_per_node: 1
       # The cluster QOS requires whole-node GPU alloc even though make_dataset is CPU-only.
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Streaming DFlash training: 2 serve replicas (TP=4) + 2 trainer nodes.
   task_1:
@@ -99,4 +99,4 @@ pipeline:
       gpus_per_node: 4
       # Pin 0.22.0: 0.22.1 regressed Kimi serve (profile_run runs the ViT and hits
       # a `fmax()` crash in kimi_k25_vit.py); 0.17.0 fails SpeculativeConfig validation.
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3.yaml
index 58efa66c1d6..532f7540ab9 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3.yaml
@@ -37,7 +37,7 @@ pipeline:
       ntasks_per_node: 1
       # The cluster QOS requires whole-node GPU alloc (4) even though make_dataset is CPU-only.
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming EAGLE3 training (node0 serve TP=4 / node1 train), 4 GPU/node
   task_1:
@@ -79,7 +79,7 @@ pipeline:
       segment: 2
       ntasks_per_node: 1
       gpus_per_node: 4
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
 
   # Step 3: Benchmark speculative decoding (VLLM backend, Kimi served at TP=4)
   task_2:
@@ -103,4 +103,4 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 4
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3_multi_node.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3_multi_node.yaml
index c43a812de7a..48009cc7cd0 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3_multi_node.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/hf_streaming_eagle3_multi_node.yaml
@@ -37,7 +37,7 @@ pipeline:
       ntasks_per_node: 1
       # The cluster QOS requires whole-node GPU allocation though make_dataset is CPU-only.
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   task_1:
     script: common/eagle3/train_eagle_streaming.sh
@@ -81,7 +81,7 @@ pipeline:
       gpus_per_node: 4
       # Pin 0.22.0: 0.22.1 regressed Kimi serve (profile_run runs the ViT and hits
       # a `fmax()` crash in kimi_k25_vit.py); 0.17.0 fails SpeculativeConfig validation.
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
 
   task_2:
     script: common/specdec_bench/quick_check.sh
@@ -104,4 +104,4 @@ pipeline:
       ntasks_per_node: 1
       gpus_per_node: 4
       # Match the serve task's pin (see task_1) — 0.22.1 broke Kimi serve.
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
diff --git a/tools/launcher/examples/moonshotai/Kimi-K2.5/specdec_bench.yaml b/tools/launcher/examples/moonshotai/Kimi-K2.5/specdec_bench.yaml
index cad5cd89df9..a25a3fe3452 100644
--- a/tools/launcher/examples/moonshotai/Kimi-K2.5/specdec_bench.yaml
+++ b/tools/launcher/examples/moonshotai/Kimi-K2.5/specdec_bench.yaml
@@ -12,7 +12,7 @@
 # `prepare_data.py --dataset speed --config all`, then replace --mtbench with
 # `--dataset speed` + `--dataset_path .../data/speed/<split>`.
 #
-# NOTE on container: pinned to the aarch64 v0.22.0 image (GB200).
+# NOTE on container: multi-arch v0.22.0 tag (resolves to amd64 on x86_64, arm64 on GB200).
 #
 # Run ON the cluster login node (paramiko can't reach the cluster through its login proxy):
 #   export SLURM_HOST=localhost SLURM_ACCOUNT=<your_account> \
@@ -58,4 +58,4 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 4
-      container: vllm/vllm-openai:v0.22.0-aarch64
+      container: vllm/vllm-openai:v0.22.0
diff --git a/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_dflash_multi_node.yaml b/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_dflash_multi_node.yaml
index f17604410fa..6d86e630738 100644
--- a/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_dflash_multi_node.yaml
+++ b/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_dflash_multi_node.yaml
@@ -46,7 +46,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming DFlash training — 2 serve replicas (TP=2) + 2 trainer nodes (2 GPU each).
   # DFlash extracts 5 target layers (build_target_layer_ids(24,5)=[1,6,11,16,21], the
diff --git a/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_eagle3_multi_node.yaml b/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_eagle3_multi_node.yaml
index 7ad003d8fd8..92885f105f6 100644
--- a/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_eagle3_multi_node.yaml
+++ b/tools/launcher/examples/openai/gpt-oss-20b/hf_streaming_eagle3_multi_node.yaml
@@ -41,7 +41,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 1
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc10
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Streaming EAGLE3 training — 2 serve replicas (TP=2) + 2 trainer nodes (2 GPU each).
   # Capture ids: default_eagle_aux_layer_ids(24)=[1,11,20] +1, plus final layer 24.
diff --git a/tools/launcher/slurm_config.py b/tools/launcher/slurm_config.py
index 758fb2bda26..516789c680f 100644
--- a/tools/launcher/slurm_config.py
+++ b/tools/launcher/slurm_config.py
@@ -71,7 +71,7 @@ def slurm_factory(
     nodes: int = 1,
     ntasks_per_node: int = 1,
     gpus_per_node: Optional[int] = 1,
-    container: str = "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc8",
+    container: str = "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20",
     modelopt_install_path: str = "/usr/local/lib/python3.12/dist-packages/modelopt",
     container_mounts: list[str] = [
         "{}:/hf-local".format(os.environ.get("SLURM_HF_LOCAL", "/hf-local")),

From 1748dbfd8b536a85dd42f11742b8849957161dd2 Mon Sep 17 00:00:00 2001
From: noeyy-mino <174223378+noeyy-mino@users.noreply.github.com>
Date: Thu, 16 Jul 2026 08:43:06 +0000
Subject: [PATCH 2/2] upgrade TRT LLM container

Signed-off-by: noeyy-mino <174223378+noeyy-mino@users.noreply.github.com>
---
 .../examples/Qwen/Qwen3-8B/eagle3_quick_check.yaml        | 8 ++++----
 .../examples/Qwen/Qwen3-8B/hf_offline_eagle3.yaml         | 6 +++---
 .../examples/Qwen/Qwen3-8B/hf_offline_eagle3_ptq.yaml     | 6 +++---
 tools/launcher/examples/Qwen/Qwen3-8B/hf_ptq.yaml         | 2 +-
 4 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/eagle3_quick_check.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/eagle3_quick_check.yaml
index 7078255eb9c..317649b55cc 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/eagle3_quick_check.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/eagle3_quick_check.yaml
@@ -39,7 +39,7 @@ pipeline:
       _factory_: "slurm_factory"
       nodes: 1
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc2
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Dump hidden states from target model
   task_1:
@@ -56,7 +56,7 @@ pipeline:
       _factory_: "slurm_factory"
       nodes: 1
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc2
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 3: Train EAGLE3 draft head (offline, single task)
   task_2:
@@ -78,7 +78,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc2
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 4: Benchmark speculative decoding (VLLM backend)
   #
@@ -90,7 +90,7 @@ pipeline:
   #   TRTLLM_LAUNCH_SCRIPT: trtllm-llmapi-launch
   # slurm_config:
   #   ntasks_per_node: 4
-  #   container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc2
+  #   container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
   #
   # To use SGLang
   #
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3.yaml
index f0a99514a10..1fe610e989e 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3.yaml
@@ -43,7 +43,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 8
       gpus_per_node: 8
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Dump hidden states from target model
   task_1:
@@ -61,7 +61,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 8
       gpus_per_node: 8
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 3: Train EAGLE3 draft head (offline, single task)
   task_2:
@@ -79,7 +79,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 8
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 4: Benchmark speculative decoding (VLLM backend)
   task_3:
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3_ptq.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3_ptq.yaml
index 58427e67ac2..c2e2dcde76e 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3_ptq.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_offline_eagle3_ptq.yaml
@@ -44,7 +44,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 4
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 2: Dump hidden states from target model
   task_1:
@@ -62,7 +62,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 4
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 3: Train EAGLE3 draft head (offline, single task) and export checkpoint
   task_2:
@@ -87,7 +87,7 @@ pipeline:
       nodes: 1
       ntasks_per_node: 1
       gpus_per_node: 4
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20
 
   # Step 4: PTQ — quantize the EAGLE3 model using offline hidden states
   task_3:
diff --git a/tools/launcher/examples/Qwen/Qwen3-8B/hf_ptq.yaml b/tools/launcher/examples/Qwen/Qwen3-8B/hf_ptq.yaml
index 55083874837..f17ca10ca98 100644
--- a/tools/launcher/examples/Qwen/Qwen3-8B/hf_ptq.yaml
+++ b/tools/launcher/examples/Qwen/Qwen3-8B/hf_ptq.yaml
@@ -48,4 +48,4 @@ pipeline:
       ntasks_per_node: 1
       gpus_per_node: 1
       time: "04:00:00"
-      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc7
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc20