Skip to content

Commit 6292847

Browse files
Oseltamivirclaude
andcommitted
bench: temporarily disable broken gb200 sglang mtp3 disagg entries
All 7 disagg entries produce 0 output tokens (dynamo multi-node prefill registration bug, see runs 25785003012 and 25812320128). Commenting them out so dispatches of this matrix key only run the agg sweep added in 8407412. Easy to restore once the dynamo bug is fixed upstream. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 parent 8407412 commit 6292847

1 file changed

Lines changed: 105 additions & 100 deletions

File tree

.github/configs/nvidia-master.yaml

Lines changed: 105 additions & 100 deletions
Original file line numberDiff line numberDiff line change
@@ -8241,106 +8241,111 @@ dsv4-fp4-gb200-dynamo-sglang-mtp3:
82418241
- isl: 8192
82428242
osl: 256
82438243
search-space:
8244-
# GB200 extrapolation of the GB300 SGLang MTP3 run 25761192586.
8245-
# TP4/DEP4 points are widened to TP8/DEP8 for GB200.
8246-
- spec-decoding: mtp
8247-
conc-list: [512]
8248-
prefill:
8249-
num-worker: 1
8250-
tp: 8
8251-
ep: 1
8252-
dp-attn: false
8253-
additional-settings:
8254-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml"
8255-
decode:
8256-
num-worker: 1
8257-
tp: 8
8258-
ep: 1
8259-
dp-attn: false
8260-
- spec-decoding: mtp
8261-
conc-list: [2048]
8262-
prefill:
8263-
num-worker: 1
8264-
tp: 8
8265-
ep: 8
8266-
dp-attn: true
8267-
additional-settings:
8268-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml"
8269-
decode:
8270-
num-worker: 6
8271-
tp: 8
8272-
ep: 1
8273-
dp-attn: false
8274-
- spec-decoding: mtp
8275-
conc-list: [4096]
8276-
prefill:
8277-
num-worker: 1
8278-
tp: 8
8279-
ep: 8
8280-
dp-attn: true
8281-
additional-settings:
8282-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep8-mtp.yaml"
8283-
decode:
8284-
num-worker: 1
8285-
tp: 8
8286-
ep: 8
8287-
dp-attn: true
8288-
- spec-decoding: mtp
8289-
conc-list: [4096]
8290-
prefill:
8291-
num-worker: 1
8292-
tp: 8
8293-
ep: 8
8294-
dp-attn: true
8295-
additional-settings:
8296-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml"
8297-
decode:
8298-
num-worker: 1
8299-
tp: 16
8300-
ep: 16
8301-
dp-attn: true
8302-
- spec-decoding: mtp
8303-
conc-list: [12288]
8304-
prefill:
8305-
num-worker: 5
8306-
tp: 8
8307-
ep: 8
8308-
dp-attn: true
8309-
additional-settings:
8310-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep8-mtp-c12288.yaml"
8311-
decode:
8312-
num-worker: 1
8313-
tp: 8
8314-
ep: 8
8315-
dp-attn: true
8316-
- spec-decoding: mtp
8317-
conc-list: [16384]
8318-
prefill:
8319-
num-worker: 6
8320-
tp: 8
8321-
ep: 8
8322-
dp-attn: true
8323-
additional-settings:
8324-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep8-mtp-c16384.yaml"
8325-
decode:
8326-
num-worker: 1
8327-
tp: 8
8328-
ep: 8
8329-
dp-attn: true
8330-
- spec-decoding: mtp
8331-
conc-list: [8192]
8332-
prefill:
8333-
num-worker: 4
8334-
tp: 8
8335-
ep: 8
8336-
dp-attn: true
8337-
additional-settings:
8338-
- "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep8-mtp-c8192.yaml"
8339-
decode:
8340-
num-worker: 1
8341-
tp: 8
8342-
ep: 8
8343-
dp-attn: true
8244+
# ===== DISAGG ENTRIES TEMPORARILY DISABLED =====
8245+
# All 7 produce 0 output tokens due to dynamo multi-node prefill
8246+
# registration bug on GB200. Restore by uncommenting once dynamo
8247+
# is fixed upstream. Agg sweep below stays active.
8248+
# # GB200 extrapolation of the GB300 SGLang MTP3 run 25761192586.
8249+
# # TP4/DEP4 points are widened to TP8/DEP8 for GB200.
8250+
# - spec-decoding: mtp
8251+
# conc-list: [512]
8252+
# prefill:
8253+
# num-worker: 1
8254+
# tp: 8
8255+
# ep: 1
8256+
# dp-attn: false
8257+
# additional-settings:
8258+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml"
8259+
# decode:
8260+
# num-worker: 1
8261+
# tp: 8
8262+
# ep: 1
8263+
# dp-attn: false
8264+
# - spec-decoding: mtp
8265+
# conc-list: [2048]
8266+
# prefill:
8267+
# num-worker: 1
8268+
# tp: 8
8269+
# ep: 8
8270+
# dp-attn: true
8271+
# additional-settings:
8272+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml"
8273+
# decode:
8274+
# num-worker: 6
8275+
# tp: 8
8276+
# ep: 1
8277+
# dp-attn: false
8278+
# - spec-decoding: mtp
8279+
# conc-list: [4096]
8280+
# prefill:
8281+
# num-worker: 1
8282+
# tp: 8
8283+
# ep: 8
8284+
# dp-attn: true
8285+
# additional-settings:
8286+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep8-mtp.yaml"
8287+
# decode:
8288+
# num-worker: 1
8289+
# tp: 8
8290+
# ep: 8
8291+
# dp-attn: true
8292+
# - spec-decoding: mtp
8293+
# conc-list: [4096]
8294+
# prefill:
8295+
# num-worker: 1
8296+
# tp: 8
8297+
# ep: 8
8298+
# dp-attn: true
8299+
# additional-settings:
8300+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml"
8301+
# decode:
8302+
# num-worker: 1
8303+
# tp: 16
8304+
# ep: 16
8305+
# dp-attn: true
8306+
# - spec-decoding: mtp
8307+
# conc-list: [12288]
8308+
# prefill:
8309+
# num-worker: 5
8310+
# tp: 8
8311+
# ep: 8
8312+
# dp-attn: true
8313+
# additional-settings:
8314+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep8-mtp-c12288.yaml"
8315+
# decode:
8316+
# num-worker: 1
8317+
# tp: 8
8318+
# ep: 8
8319+
# dp-attn: true
8320+
# - spec-decoding: mtp
8321+
# conc-list: [16384]
8322+
# prefill:
8323+
# num-worker: 6
8324+
# tp: 8
8325+
# ep: 8
8326+
# dp-attn: true
8327+
# additional-settings:
8328+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep8-mtp-c16384.yaml"
8329+
# decode:
8330+
# num-worker: 1
8331+
# tp: 8
8332+
# ep: 8
8333+
# dp-attn: true
8334+
# - spec-decoding: mtp
8335+
# conc-list: [8192]
8336+
# prefill:
8337+
# num-worker: 4
8338+
# tp: 8
8339+
# ep: 8
8340+
# dp-attn: true
8341+
# additional-settings:
8342+
# - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep8-mtp-c8192.yaml"
8343+
# decode:
8344+
# num-worker: 1
8345+
# tp: 8
8346+
# ep: 8
8347+
# dp-attn: true
8348+
# ===== END DISABLED BLOCK =====
83448349

83458350
# Aggregated fallback sweep: single agg worker, TP=8 across 2 nodes,
83468351
# DP attention + EAGLE MTP. Sidesteps the dynamo multi-node prefill

0 commit comments

Comments
 (0)