[OMNIML-4962] specdec_bench parent — Qwen/Qwen3.5-4B / DFlash / vLLM

tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench_dflash_vllm.yaml

@@ -0,0 +1,99 @@
+# SPEED-bench DFlash speculative-decoding run for Qwen3.5-4B via vLLM.
+#
+# Companion to specdec_bench_mtp_vllm.yaml. This variant exercises the
+# z-lab/Qwen3.5-4B-DFlash external draft model. DFlash uses
+# vLLM's `speculative_num_draft_tokens`, which the specdec_bench wrapper
+# computes from `--block_size`. `--block_size = draft_length + 1`
+# (e.g. draft_length=3 → block_size=4; draft_length=7 → block_size=8).
+#
+# Container: `vllm/vllm-openai:v0.22.1`+ is required. The `dflash`
+# speculative method landed in vLLM v0.22.0; older tags reject
+# `--speculative_algorithm DFLASH` with "Input should be 'ngram', ...,
+# 'mtp'".
+#
+# Cells in OMNIML-4961 (t0_d3 / t0_d7 / t1_d3 / t1_d7) override per-cell
+# knobs (temperature, block_size, save_dir, num_requests) via
+# pipeline.task_N.args+=[...] at slurm-invoke time — no per-cell file is
+# committed (post-#1564 convention).
+#
+# Slurm run on cw_dfw:
+#   uv run slurm.py \
+#     --yaml modules/Model-Optimizer/tools/launcher/examples/Qwen/Qwen3.5-4B/specdec_bench_dflash_vllm.yaml \
+#     --yes detach=true \
+#     pipeline.task_0.args+=["--temperature 0","--block_size 4","--save_dir /scratchspace/<sweep>/qualitative"] \
+#     pipeline.task_1.args+=["--temperature 0","--block_size 4","--save_dir /scratchspace/<sweep>/throughput_32k","--num_requests 80","--max_seq_len 65536"]
+#
+# Reference run: cicd_1780697785 (OMNIML-4962, cw_dfw) produced
+#   qualitative   Average_AL = 5.0395
+#   throughput_32k Average_AL = 6.8958
+# See OMNIML-4962 INTERN-ARTIFACTS worklog for the full payload.
+
+job_name: Qwen3.5-4B_specdec_bench_dflash_vllm
+
+pipeline:
+  global_vars:
+    hf_model: /hf-local/Qwen/Qwen3.5-4B
+    draft_model: /hf-local/z-lab/Qwen3.5-4B-DFlash
+
+  # task_0: SPEED qualitative split — quality / acceptance-rate numbers.
+  # tp_size=2 + concurrency=32 trades aa_timing fidelity for ~30x
+  # wall-clock speedup; acceptance-length (AL) is concurrency-independent
+  # and is the primary metric we care about for this split.
+  task_0:
+    script: common/specdec_bench/run.sh
+    args:
+      - --dataset speed
+      - --dataset_path /hf-local/nvidia/SPEED-Bench-Internal/qualitative
+      - --engine VLLM
+      - --speculative_algorithm DFLASH
+      - --draft_model_dir <<global_vars.draft_model>>
+      - --block_size 4
+      - --tp_size 2
+      - --ep_size 1
+      - --concurrency 32
+      - --output_length 4096
+      - --aa_timing
+      - --show_progress
+      - --save_dir /scratchspace/{sweep_name_default}/qualitative
+    environment:
+      - HF_MODEL_CKPT: <<global_vars.hf_model>>
+      - HF_LOCAL: /hf-local
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 1
+      gpus_per_node: 2
+      container: vllm/vllm-openai:v0.22.1
+
+  # task_1: SPEED throughput_32k split — long-context throughput.
+  # `--num_requests 80` caps the run at 80 samples (split has 1,536) so
+  # it fits in the 4h Slurm time-limit; each 32K-input sample takes
+  # ~60-90s. tp_size=2 doubles the KV-cache budget across 2 GPUs;
+  # concurrency=8 keeps 8 * 32K = 256K tokens of in-flight KV under that
+  # doubled budget.
+  task_1:
+    script: common/specdec_bench/run.sh
+    args:
+      - --dataset speed
+      - --dataset_path /hf-local/nvidia/SPEED-Bench-Internal/throughput_32k
+      - --engine VLLM
+      - --speculative_algorithm DFLASH
+      - --draft_model_dir <<global_vars.draft_model>>
+      - --block_size 4
+      - --tp_size 2
+      - --ep_size 1
+      - --concurrency 8
+      - --num_requests 80
+      - --output_length 4096
+      - --aa_timing
+      - --show_progress
+      - --save_dir /scratchspace/{sweep_name_default}/throughput_32k
+    environment:
+      - HF_MODEL_CKPT: <<global_vars.hf_model>>
+      - HF_LOCAL: /hf-local
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 1
+      gpus_per_node: 2
+      container: vllm/vllm-openai:v0.22.1