Skip to content
Draft
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 40 additions & 3 deletions configs/nvidia-master.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -6699,9 +6699,12 @@ qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
disagg: true
scenarios:
agentic-coding:
- search-space:
- dram-utilization: 0.60953
search-space:
# 1P1D: 1 prefill (TP1/EP1/dp-attn), 1 decode (TP2/EP2), conc=44
- spec-decoding: "mtp"
kv-offloading: dram
kv-offload-backend: { name: native, version: "1.3.0rc24" }
conc-list: [44]
prefill:
num-worker: 1
Expand All @@ -6715,8 +6718,12 @@ qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
tp: 2
ep: 2
dp-attn: false
- dram-utilization: 0.15239
search-space:
# 1P7D: 1 prefill (TP4/EP4/dp-attn), 7 decode (TP8/EP8), conc=7
- spec-decoding: "mtp"
kv-offloading: dram
kv-offload-backend: { name: native, version: "1.3.0rc24" }
conc-list: [7]
prefill:
num-worker: 1
Expand All @@ -6730,8 +6737,12 @@ qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
tp: 8
ep: 8
dp-attn: false
- dram-utilization: 0.60953
search-space:
# 2P2D: 2 prefill (TP1/EP1/dp-attn), 2 decode (TP2/EP2), conc=52
- spec-decoding: "mtp"
kv-offloading: dram
kv-offload-backend: { name: native, version: "1.3.0rc24" }
conc-list: [52]
prefill:
num-worker: 2
Expand All @@ -6745,8 +6756,12 @@ qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
tp: 2
ep: 2
dp-attn: false
- dram-utilization: 0.30477
search-space:
# 2P3D: 2 prefill (TP2/EP2), 3 decode (TP8/EP8), conc=96
- spec-decoding: "mtp"
kv-offloading: dram
kv-offload-backend: { name: native, version: "1.3.0rc24" }
conc-list: [96]
prefill:
num-worker: 2
Expand All @@ -6760,8 +6775,12 @@ qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
tp: 8
ep: 8
dp-attn: false
- dram-utilization: 0.15239
search-space:
# 3P1D: 3 prefill (TP4/EP4/dp-attn), 1 decode (TP16/EP16/dp-attn), conc=565
- spec-decoding: "mtp"
kv-offloading: dram
kv-offload-backend: { name: native, version: "1.3.0rc24" }
conc-list: [565]
prefill:
num-worker: 3
Expand All @@ -6777,6 +6796,8 @@ qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
dp-attn: true
# 3P2D: 3 prefill (TP4/EP4/dp-attn), 2 decode (TP4/EP4/dp-attn), conc=704
- spec-decoding: "mtp"
kv-offloading: dram
kv-offload-backend: { name: native, version: "1.3.0rc24" }
conc-list: [704]
prefill:
num-worker: 3
Expand Down Expand Up @@ -7646,10 +7667,13 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
disagg: true
scenarios:
agentic-coding:
- search-space:
- dram-utilization: 0.19957
search-space:
# Keep the checked-in recipes on real MTP verification. Throughput jobs
# inject the committed golden AL at launch; EVAL_ONLY leaves them real.
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [1152]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand All @@ -7667,6 +7691,8 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
ep: 12
dp-attn: true
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [1024]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand All @@ -7684,6 +7710,8 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
ep: 16
dp-attn: true
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [256]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand Down Expand Up @@ -7788,10 +7816,13 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-disagg:
disagg: true
scenarios:
agentic-coding:
- search-space:
- dram-utilization: 0.15523
search-space:
# Keep the checked-in recipes on real MTP verification. Throughput jobs
# inject the committed golden AL at launch; EVAL_ONLY leaves them real.
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [576]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand All @@ -7809,6 +7840,8 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-disagg:
ep: 12
dp-attn: true
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [512]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand All @@ -7826,6 +7859,8 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-disagg:
ep: 16
dp-attn: true
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [256]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand All @@ -7843,6 +7878,8 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-disagg:
ep: 8
dp-attn: true
- spec-decoding: mtp
kv-offloading: dram
kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
conc-list: [128]
router: { name: dynamo-router, version: "1.3.0.dev20260720" }
prefill:
Expand Down