sglang: add fp8 8k1k and fp4 1k1k (#274)

* go

* typo

* typo...

* more

Author: ishandhanani

.github/configs/nvidia-master.yaml

@@ -798,6 +798,7 @@ dsr1-fp8-gb200-dynamo-sglang:
         additional-settings:
         - "PREFILL_NODES=4"
         - "N_ADDITIONAL_FRONTENDS=9"
+        - "SCRIPT_MODE=max-tpt"
       decode:
         num-worker: 1
         tp: 1
@@ -852,22 +853,112 @@ dsr1-fp8-gb200-dynamo-sglang:
   - isl: 8192
     osl: 1024
     search-space:
+    # Low latency (1 prefill worker at DEP4 and 1 decode worker at DEP4)
     - spec-decoding: "none"
-      conc-list: [ 128, 256, 384, 448, 512, 576, 1024, 2048, 4096 ]
+      conc-list: [ 4, 8, 16, 32, 64, 128, 256, 512 ]
       prefill:
-        num-worker: 6
-        # tp, ep, and dp-attn do nothing because they are hardcoded in the following file:
-        # https://github.com/Elnifio/dynamo/blob/update-result-file-name/components/backends/sglang/slurm_jobs/scripts/gb200-fp8.sh
+        num-worker: 1
         tp: 1
         ep: 1
         dp-attn: true
         additional-settings:
-        - "PREFILL_NODES=12"
+        - "PREFILL_NODES=1"
         - "N_ADDITIONAL_FRONTENDS=8"
       decode:
         num-worker: 1
         tp: 1
         ep: 1
         dp-attn: true
         additional-settings:
-        - "DECODE_NODES=6"
+        - "DECODE_NODES=1"
+
+    # Middle and top of curve (5 prefill workers each at DEP8 and 1 decode worker at DEP32)
+    - spec-decoding: "none"
+      conc-list: [ 512, 1024, 2048, 6144 ]
+      prefill:
+        num-worker: 5
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "PREFILL_NODES=2"
+        - "N_ADDITIONAL_FRONTENDS=8"
+      decode:
+        num-worker: 1
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "DECODE_NODES=8"
+
+dsr1-fp4-gb200-dynamo-sglang:
+  # TODO: swap
+  image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:0.5.1-rc0.pre1
+  # TODO: what is the right name?
+  model: deepseek-ai/DeepSeek-R1-0528-fp4-v2
+  model-prefix: dsr1
+  runner: gb200
+  precision: fp4
+  framework: dynamo-sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    # Low latency (1 prefill worker at DEP4 and 2 decode workers at DEP4)
+    - spec-decoding: "none"
+      conc-list: [ 4, 8, 32, 64, 128, 112, 128, 256 ]
+      prefill:
+        num-worker: 1
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "N_ADDITIONAL_FRONTENDS=8"
+      decode:
+        num-worker: 2
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "DECODE_NODES=2"
+
+    # Mid curve (1 prefill worker at DEP4 and 1 decode workers at DEP48)
+    - spec-decoding: "none"
+      conc-list: [ 512, 1024, 2048, 4096, 8192 ]
+      prefill:
+        num-worker: 1
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "N_ADDITIONAL_FRONTENDS=8"
+      decode:
+        num-worker: 2
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "DECODE_NODES=12"
+
+    # Top of curve (1 prefill worker at DEP4 and 1 decode worker at DEP32)
+    - spec-decoding: "none"
+      conc-list: [ 8192, 12000, 15000 ]
+      prefill:
+        num-worker: 1
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "N_ADDITIONAL_FRONTENDS=8"
+      decode:
+        num-worker: 2
+        tp: 1
+        ep: 1
+        dp-attn: true
+        additional-settings:
+        - "DECODE_NODES=8"

benchmarks/dsr1_fp4_gb200_dynamo-sglang_slurm.sh

@@ -0,0 +1,37 @@
+
+#!/bin/bash
+
+set -x
+
+source "$(dirname "$0")/benchmark_lib.sh"
+
+check_env_vars CONC_LIST ISL OSL IMAGE SPEC_DECODING MODEL_PATH \
+    PREFILL_NUM_WORKERS PREFILL_TP PREFILL_EP PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS DECODE_TP DECODE_EP DECODE_DP_ATTN \
+    PREFILL_NODES DECODE_NODES N_ADDITIONAL_FRONTENDS SGL_SLURM_JOBS_PATH # SGL_SLURM_JOBS_PATH FIXME
+
+# Always clone and setup Dynamo
+echo "Cloning Dynamo repository..."
+git clone --branch ishan/sa-1.1-sgl-dsr1 https://github.com/ai-dynamo/dynamo.git
+
+cd "$SGL_SLURM_JOBS_PATH"
+
+# Set up SGL launch script-specific environment variables
+export TIME_LIMIT="04:00:00"
+export MODEL_PATH=$MODEL_PATH
+export CONFIG_DIR=$CONFIG_DIR
+export CONTAINER_IMAGE=$IMAGE
+export GPU_TYPE="gb200-fp4"
+
+# Launch jobs based on ISL/OSL
+# Replace ' ' in CONC_LIST with 'x' such that the concurrency list is represented
+# by a list of numbers delimted by 'x'. This is because of how the underlying launch script
+# expects the concurrencies.
+bash ./submit_disagg.sh $PREFILL_NODES \
+    $PREFILL_NUM_WORKERS \
+    $DECODE_NODES \
+    $DECODE_NUM_WORKERS \
+    $N_ADDITIONAL_FRONTENDS \
+    $ISL $OSL "${CONC_LIST// /x}" inf \
+    $GPU_TYPE \
+    $SCRIPT_MODE

benchmarks/dsr1_fp8_gb200_dynamo-sglang_slurm.sh

@@ -12,11 +12,7 @@ check_env_vars CONC_LIST ISL OSL IMAGE SPEC_DECODING MODEL_PATH \
 
 # Always clone and setup Dynamo
 echo "Cloning Dynamo repository..."
-if [ "$ISL" = "1024" ] && [ "$OSL" = "1024" ]; then
-    git clone --branch ishan/sa-1.1-sgl-dsr1-fp8 https://github.com/ai-dynamo/dynamo.git
-else
-    git clone --branch update-result-file-name https://github.com/Elnifio/dynamo.git
-fi
+git clone --branch ishan/sa-1.1-sgl-dsr1 https://github.com/ai-dynamo/dynamo.git
 
 cd "$SGL_SLURM_JOBS_PATH"
 
@@ -25,6 +21,7 @@ export TIME_LIMIT="04:00:00"
 export MODEL_PATH=$MODEL_PATH
 export CONFIG_DIR=$CONFIG_DIR
 export CONTAINER_IMAGE=$IMAGE
+export GPU_TYPE="gb200-fp8"
 
 # Launch jobs based on ISL/OSL
 # Replace ' ' in CONC_LIST with 'x' such that the concurrency list is represented
@@ -36,4 +33,5 @@ bash ./submit_disagg.sh $PREFILL_NODES \
     $DECODE_NUM_WORKERS \
     $N_ADDITIONAL_FRONTENDS \
     $ISL $OSL "${CONC_LIST// /x}" inf \
+    $GPU_TYPE \
     $SCRIPT_MODE