[AMD][DI][CI] 2/N Add DSV4 DP8/EP8 and MTP MI355X 1P1D nightly recipes (#29784)
Co-authored-by: bingxche <bingxche@amd.com>
This commit is contained in:
@@ -117,3 +117,206 @@ dsv4pro-fp4-mi355x-sglang:
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d.yaml
|
||||
|
||||
# AMD MI355X 2-node 1P1D disaggregation over MORI with extra topology / MTP
|
||||
# coverage on the SAME four DeepSeek-V4 model x precision combos as the base
|
||||
# MORI blocks above (which stay TP8, no MTP). Three variants per model:
|
||||
# * -mtp : TP8 + EAGLE MTP (recipe `mtp.enabled`)
|
||||
# * -dp8ep8 : DP-attention 8 + narrow within-node EP8
|
||||
# * -dp8ep8-mtp : DP8 + narrow EP8 + EAGLE MTP
|
||||
# launch_mi355x.sh reads expert-/data-parallel-size and the `mtp:` section to
|
||||
# append --ep-size / --enable-dp-attention --dp-size / --speculative-* flags.
|
||||
|
||||
dsv4flash-fp8-mi355x-mtp-sglang:
|
||||
model: sgl-project/DeepSeek-V4-Flash-FP8
|
||||
model-prefix: dsv4flash
|
||||
model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
|
||||
runner: mi355x
|
||||
precision: fp8
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-mtp.yaml
|
||||
|
||||
dsv4flash-fp8-mi355x-dp8ep8-sglang:
|
||||
model: sgl-project/DeepSeek-V4-Flash-FP8
|
||||
model-prefix: dsv4flash
|
||||
model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
|
||||
runner: mi355x
|
||||
precision: fp8
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-dp8ep8.yaml
|
||||
|
||||
dsv4flash-fp8-mi355x-dp8ep8-mtp-sglang:
|
||||
model: sgl-project/DeepSeek-V4-Flash-FP8
|
||||
model-prefix: dsv4flash
|
||||
model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
|
||||
runner: mi355x
|
||||
precision: fp8
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-dp8ep8-mtp.yaml
|
||||
|
||||
dsv4pro-fp8-mi355x-mtp-sglang:
|
||||
model: sgl-project/DeepSeek-V4-Pro-FP8
|
||||
model-prefix: dsv4pro
|
||||
model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
|
||||
runner: mi355x
|
||||
precision: fp8
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
# conc256 excluded: disagg-decode SWA hybrid pool retract->get_cpu_copy
|
||||
# is an upstream NotImplementedError (crashes decode). See recipe.
|
||||
- conc-list: [1, 8, 16, 32, 64, 128]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/1p1d-mtp.yaml
|
||||
|
||||
dsv4pro-fp8-mi355x-dp8ep8-sglang:
|
||||
model: sgl-project/DeepSeek-V4-Pro-FP8
|
||||
model-prefix: dsv4pro
|
||||
model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
|
||||
runner: mi355x
|
||||
precision: fp8
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/1p1d-dp8ep8.yaml
|
||||
|
||||
dsv4pro-fp8-mi355x-dp8ep8-mtp-sglang:
|
||||
model: sgl-project/DeepSeek-V4-Pro-FP8
|
||||
model-prefix: dsv4pro
|
||||
model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
|
||||
runner: mi355x
|
||||
precision: fp8
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/1p1d-dp8ep8-mtp.yaml
|
||||
|
||||
dsv4flash-fp4-mi355x-mtp-sglang:
|
||||
model: deepseek-ai/DeepSeek-V4-Flash
|
||||
model-prefix: dsv4flash
|
||||
model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
|
||||
runner: mi355x
|
||||
precision: fp4
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mtp.yaml
|
||||
|
||||
dsv4flash-fp4-mi355x-dp8ep8-sglang:
|
||||
model: deepseek-ai/DeepSeek-V4-Flash
|
||||
model-prefix: dsv4flash
|
||||
model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
|
||||
runner: mi355x
|
||||
precision: fp4
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-dp8ep8.yaml
|
||||
|
||||
dsv4flash-fp4-mi355x-dp8ep8-mtp-sglang:
|
||||
model: deepseek-ai/DeepSeek-V4-Flash
|
||||
model-prefix: dsv4flash
|
||||
model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
|
||||
runner: mi355x
|
||||
precision: fp4
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-dp8ep8-mtp.yaml
|
||||
|
||||
dsv4pro-fp4-mi355x-mtp-sglang:
|
||||
model: deepseek-ai/DeepSeek-V4-Pro
|
||||
model-prefix: dsv4pro
|
||||
model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
|
||||
runner: mi355x
|
||||
precision: fp4
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d-mtp.yaml
|
||||
|
||||
dsv4pro-fp4-mi355x-dp8ep8-sglang:
|
||||
model: deepseek-ai/DeepSeek-V4-Pro
|
||||
model-prefix: dsv4pro
|
||||
model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
|
||||
runner: mi355x
|
||||
precision: fp4
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d-dp8ep8.yaml
|
||||
|
||||
dsv4pro-fp4-mi355x-dp8ep8-mtp-sglang:
|
||||
model: deepseek-ai/DeepSeek-V4-Pro
|
||||
model-prefix: dsv4pro
|
||||
model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
|
||||
runner: mi355x
|
||||
precision: fp4
|
||||
framework: sglang
|
||||
multinode: true
|
||||
disagg: true
|
||||
seq-len-configs:
|
||||
- isl: 1024
|
||||
osl: 1024
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d-dp8ep8-mtp.yaml
|
||||
|
||||
Reference in New Issue
Block a user