[OMNIML-4333] training_support

tools/launcher/examples/Qwen/qwen3-v0339a-demo/step1_synth.yaml

@@ -0,0 +1,40 @@
+# EAGLE3 speculative decoding pipeline — Step 1: Data synthesis
+#
+# Queries TRT-LLM server to generate synthetic prompt samples for EAGLE3 training.
+#
+# Usage:
+#   uv run launch.py --yaml examples/Qwen/qwen3-v0339a-demo/step1_synth.yaml --yes
+#   uv run slurm.py --yaml modules/Model-Optimizer/tools/launcher/examples/Qwen/qwen3-v0339a-demo/step1_synth.yaml --yes
+
+job_name: qwen3-v0339a-demo_EAGLE3_step1_synth
+pipeline:
+  allow_to_fail: false
+  skip: false
+  note: Synthetic data generation via TRT-LLM server
+
+  global_vars:
+    hf_model: /hf-local/Qwen/qwen3-v0339a-demo
+
+  # Step 1: Data synthesis via TRT-LLM server
+  # Args before "--" go to trtllm-serve; args after "--" go to tools/query.py.
+  task_0:
+    script: common/tensorrt_llm/query.sh
+    args:
+      - --model <<global_vars.hf_model>>
+      - --tp_size 8
+      - --ep_size 8
+      - --max_num_tokens 32000
+      - --port 8000
+      - --host 0.0.0.0
+      - --trust_remote_code
+      - --
+      - --data /hf-local/modelopt/Speculative-Decoding-Prompt-Samples
+      - --save /scratchspace/data
+    environment:
+      - HF_LOCAL: /hf-local
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 8
+      gpus_per_node: 8
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0

tools/launcher/examples/Qwen/qwen3-v0339a-demo/step2_hidden.yaml

@@ -0,0 +1,34 @@
+# EAGLE3 speculative decoding pipeline — Step 2: Hidden state dump
+#
+# Runs target model to capture hidden states for offline EAGLE3 training.
+#
+# Usage:
+#   uv run launch.py --yaml examples/Qwen/qwen3-v0339a-demo/step2_hidden.yaml --yes
+#   uv run slurm.py --yaml modules/Model-Optimizer/tools/launcher/examples/Qwen/qwen3-v0339a-demo/step2_hidden.yaml --yes
+
+job_name: qwen3-v0339a-demo_EAGLE3_step2_hidden
+pipeline:
+  allow_to_fail: false
+  skip: false
+  note: Hidden state dump from target model
+
+  global_vars:
+    hf_model: /hf-local/Qwen/qwen3-v0339a-demo
+
+  # Step 2: Dump hidden states from target model
+  task_0:
+    script: common/eagle3/dump_offline_data.sh
+    args:
+      - --input-data /scratchspace/data
+      - --output-dir /scratchspace/offline_hidden_states
+      - --max-seq-len 8192
+      - --tp 8
+      - --moe-ep 8
+    environment:
+      - HF_MODEL_CKPT: <<global_vars.hf_model>>
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 8
+      gpus_per_node: 8
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0

tools/launcher/examples/Qwen/qwen3-v0339a-demo/step3_train.yaml

@@ -0,0 +1,34 @@
+# EAGLE3 speculative decoding pipeline — Step 3: Offline training
+#
+# Trains EAGLE3 draft head using collected hidden states from target model.
+#
+# Usage:
+#   uv run launch.py --yaml examples/Qwen/qwen3-v0339a-demo/step3_train.yaml --yes
+#   uv run slurm.py --yaml modules/Model-Optimizer/tools/launcher/examples/Qwen/qwen3-v0339a-demo/step3_train.yaml --yes
+
+job_name: qwen3-v0339a-demo_EAGLE3_step3_train
+pipeline:
+  allow_to_fail: false
+  skip: false
+  note: Offline EAGLE3 training
+
+  global_vars:
+    hf_model: /hf-local/Qwen/qwen3-v0339a-demo
+
+  # Step 3: Train EAGLE3 draft head (offline, single task)
+  task_0:
+    script: common/eagle3/train_eagle.sh
+    args:
+      - --config modules/Model-Optimizer/modelopt_recipes/general/speculative_decoding/eagle3.yaml
+      - model.model_name_or_path=<<global_vars.hf_model>>
+      - data.offline_data_path=/scratchspace/offline_hidden_states
+      - training.output_dir=/scratchspace/eagle3
+      - training.training_seq_len=4096
+      - training.disable_tqdm=true
+      - training.ar_validate_steps=500000
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 1
+      gpus_per_node: 8
+      container: nvcr.io/nvidia/tensorrt-llm/release:1.2.0

tools/launcher/examples/Qwen/qwen3-v0339a-demo/step4_speed_eval.yaml

@@ -0,0 +1,38 @@
+# EAGLE3 speculative decoding pipeline — Step 4: Speed evaluation
+#
+# Benchmarks speculative decoding speedup via vLLM with trained EAGLE3 draft head.
+#
+# Usage:
+#   uv run launch.py --yaml examples/Qwen/qwen3-v0339a-demo/step4_speed_eval.yaml --yes
+#   uv run slurm.py --yaml modules/Model-Optimizer/tools/launcher/examples/Qwen/qwen3-v0339a-demo/step4_speed_eval.yaml --yes
+
+job_name: qwen3-v0339a-demo_EAGLE3_step4_speed_eval
+pipeline:
+  allow_to_fail: false
+  skip: false
+  note: Speculative decoding benchmark evaluation
+
+  global_vars:
+    hf_model: /hf-local/Qwen/qwen3-v0339a-demo
+
+  # Step 4: Benchmark speculative decoding (VLLM backend)
+  task_0:
+    script: common/specdec_bench/quick_check.sh
+    args:
+      - --draft_model_dir /scratchspace/export
+      - --draft_length 3
+      - --output_length 4096
+      - --engine VLLM
+      - --tp_size 8
+      - --ep_size 1
+      - --speculative_algorithm EAGLE3
+      - --mtbench /hf-local/HuggingFaceH4/mt_bench_prompts/raw/question.jsonl
+      - --concurrency 1
+    environment:
+      - HF_MODEL_CKPT: <<global_vars.hf_model>>
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 1
+      gpus_per_node: 8
+      container: vllm/vllm-openai:latest