[AMD][DI][CI] 8/N Add GLM-5.2 MXFP4 1P1D DI/CI recipes (base + MTP + DP8/EP8) (#32120)
This commit is contained in:
@@ -532,3 +532,75 @@ kimik26-mxfp4-mi355x-ep16-mtp-sglang:
|
||||
search-space:
|
||||
- conc-list: [1, 8, 16, 32, 64, 128, 256]
|
||||
config_file: scripts/ci/slurm/recipes/mi355x-fp8/kimik26/1k1k/2p1d-ep16-mtp-mxfp4.yaml
|
||||
|
||||
# AMD MI355X 2-node 1P1D disaggregation for GLM-5.2 (MXFP4) over MORI. GLM uses
|
||||
# GlmMoeDsaForCausalLM (DeepSeek Sparse Attention, auto-selected) with a built-in
|
||||
# NextN MTP head. Four variants mirroring the DeepSeek-V4 basic tier:
|
||||
# * (base) : TP8, no MTP
|
||||
# * -mtp : TP8 + NextN MTP
|
||||
# * -dp8ep8 : DP-attention 8 + narrow within-node EP8
|
||||
# * -dp8ep8-mtp : DP8 + narrow EP8 + NextN MTP
|
||||
# NOTE: checkpoint on disk is amd/GLM-5.1-MXFP4 (only GLM MXFP4 build mirrored).
|
||||
glm52-fp4-mi355x-sglang:
|
||||
model: amd/GLM-5.2-MXFP4
|
||||
model-prefix: glm52
|
||||
model_path: /it-share/model_coverage/models--amd--GLM-5.1-MXFP4
|
||||
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/glm52/1k1k/1p1d.yaml
|
||||
|
||||
glm52-fp4-mi355x-mtp-sglang:
|
||||
model: amd/GLM-5.2-MXFP4
|
||||
model-prefix: glm52
|
||||
model_path: /it-share/model_coverage/models--amd--GLM-5.1-MXFP4
|
||||
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/glm52/1k1k/1p1d-mtp.yaml
|
||||
|
||||
glm52-fp4-mi355x-dp8ep8-sglang:
|
||||
model: amd/GLM-5.2-MXFP4
|
||||
model-prefix: glm52
|
||||
model_path: /it-share/model_coverage/models--amd--GLM-5.1-MXFP4
|
||||
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/glm52/1k1k/1p1d-dp8ep8.yaml
|
||||
|
||||
glm52-fp4-mi355x-dp8ep8-mtp-sglang:
|
||||
model: amd/GLM-5.2-MXFP4
|
||||
model-prefix: glm52
|
||||
model_path: /it-share/model_coverage/models--amd--GLM-5.1-MXFP4
|
||||
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/glm52/1k1k/1p1d-dp8ep8-mtp.yaml
|
||||
|
||||
Reference in New Issue
Block a user