[mem_cache] Clean up SWA/Mamba radix cache leftovers and drop SGLANG_ENABLE_UNIFIED_RADIX_TREE (#40780)

This commit is contained in:
Liangsheng Yin
2026-09-22 15:40:57 -07:00
committed by GitHub
parent 19ee4b566b
commit d34f7b2326
37 changed files with 107 additions and 408 deletions
@@ -739,7 +739,6 @@ Pending update...
**TP8, concurrency 1–8:**
```bash Command
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
python3 -m sglang.launch_server \
--model-path deepseek-ai/DeepSeek-V4-Pro-0813 \
--trust-remote-code \
@@ -765,7 +764,6 @@ DSv4 HiCache sizes the host tier with `--hicache-ratio` (host/device token ratio
**DEP8 (DP Attention), concurrency 64–160:**
```bash Command
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320 \
python3 -m sglang.launch_server \
--model-path deepseek-ai/DeepSeek-V4-Pro-0813 \
@@ -883,7 +881,6 @@ DeepSeek-V4-Pro-0813 bundles the DSpark head, so `--speculative-draft-model-path
**TP8, concurrency 1–16:**
```bash Command
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
@@ -922,7 +919,6 @@ Concurrency 32–48 keeps the same TP-only command and adds HiCache DRAM offload
**DP8 (DP Attention), concurrency 128–256:**
```bash Command
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
@@ -13,7 +13,7 @@ Session references are soft protection, not memory pins. Referenced KV can still
This feature is implemented only by `UnifiedRadixCache`.
```bash Command
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 python3 -m sglang.launch_server \
python3 -m sglang.launch_server \
--model-path MODEL_PATH \
--enable-session-radix-cache
```
@@ -1414,11 +1414,6 @@ SGLang supports various environment variables that can be used to configure its
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Multiplier applied to the sliding-window-attention eviction interval.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>1.0</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_ENABLE_UNIFIED_RADIX_TREE</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Use the unified radix-tree cache implementation.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>false</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_RADIX_FORCE_MISS</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Force radix-cache misses (debugging/benchmarking).</td>
@@ -209,9 +209,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -236,9 +234,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -262,9 +258,7 @@ export const config = {
},
{
match: { hw: "gb200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -289,9 +283,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -320,9 +312,7 @@ export const config = {
{
match: { hw: "h200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -352,9 +342,7 @@ export const config = {
{
match: { hw: "dgx-spark", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "multi-2" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -384,7 +372,6 @@ export const config = {
verified: true,
env: [
"SGLANG_USE_AITER=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -410,7 +397,6 @@ export const config = {
verified: true,
env: [
"SGLANG_USE_AITER=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -442,9 +428,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -470,9 +454,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -497,9 +479,7 @@ export const config = {
},
{
match: { hw: "gb200", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -525,9 +505,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -560,9 +538,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -593,9 +569,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -625,9 +599,7 @@ export const config = {
},
{
match: { hw: "gb200", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -658,9 +630,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -691,9 +661,7 @@ export const config = {
{
match: { hw: "h200", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -729,9 +697,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "dspark", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -764,9 +730,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "dspark", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -799,9 +763,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "dspark", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -834,9 +796,7 @@ export const config = {
{
match: { hw: "h200", variant: "default", quant: "nvfp4", strategy: "dspark", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -879,7 +839,6 @@ export const config = {
"NCCL_MNNVL_ENABLE=1",
"NCCL_NVLS_ENABLE=1",
"NCCL_CUMEM_ENABLE=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -908,7 +867,6 @@ export const config = {
"NCCL_MNNVL_ENABLE=1",
"NCCL_NVLS_ENABLE=1",
"NCCL_CUMEM_ENABLE=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -940,9 +898,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "bf16", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -967,9 +923,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "bf16", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -997,9 +951,7 @@ export const config = {
},
{
match: { hw: "b200", variant: "default", quant: "bf16", strategy: "balanced", nodes: "multi-2" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -1021,9 +973,7 @@ export const config = {
},
{
match: { hw: "b200", variant: "default", quant: "bf16", strategy: "mtp", nodes: "multi-2" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -1058,7 +1008,6 @@ export const config = {
{
match: { hw: "b200", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1092,7 +1041,6 @@ export const config = {
{
match: { hw: "b300", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1126,7 +1074,6 @@ export const config = {
{
match: { hw: "gb200", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1161,7 +1108,6 @@ export const config = {
match: { hw: "gb300", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1195,7 +1141,6 @@ export const config = {
match: { hw: "h200", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
"SGLANG_OPT_USE_INKLING_SHEARED_BIAS=0",
@@ -1231,7 +1176,6 @@ export const config = {
match: { hw: "gb300", variant: "lora", quant: "bf16", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
"SGLANG_OPT_USE_JIT_KERNEL_MOE_ALIGN=1",
@@ -1264,7 +1208,6 @@ export const config = {
match: { hw: "h200", variant: "lora", quant: "bf16", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
"SGLANG_OPT_USE_JIT_KERNEL_MOE_ALIGN=1",
@@ -206,9 +206,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -233,9 +231,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -259,9 +255,7 @@ export const config = {
},
{
match: { hw: "gb200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -286,9 +280,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -317,9 +309,7 @@ export const config = {
{
match: { hw: "h200", variant: "default", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -349,7 +339,6 @@ export const config = {
verified: true,
env: [
"SGLANG_USE_AITER=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -375,7 +364,6 @@ export const config = {
verified: true,
env: [
"SGLANG_USE_AITER=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -407,9 +395,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -435,9 +421,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -462,9 +446,7 @@ export const config = {
},
{
match: { hw: "gb200", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -490,9 +472,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "long_context", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -525,9 +505,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -558,9 +536,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -590,9 +566,7 @@ export const config = {
},
{
match: { hw: "gb200", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -623,9 +597,7 @@ export const config = {
{
match: { hw: "gb300", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -656,9 +628,7 @@ export const config = {
{
match: { hw: "h200", variant: "default", quant: "nvfp4", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -696,9 +666,7 @@ export const config = {
{
match: { hw: "b200", variant: "default", quant: "nvfp4", strategy: "dspark", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -742,7 +710,6 @@ export const config = {
"NCCL_MNNVL_ENABLE=1",
"NCCL_NVLS_ENABLE=1",
"NCCL_CUMEM_ENABLE=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -772,7 +739,6 @@ export const config = {
"NCCL_MNNVL_ENABLE=1",
"NCCL_NVLS_ENABLE=1",
"NCCL_CUMEM_ENABLE=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
flags: [
"--trust-remote-code",
@@ -804,9 +770,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "bf16", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -831,9 +795,7 @@ export const config = {
{
match: { hw: "b300", variant: "default", quant: "bf16", strategy: "mtp", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -863,9 +825,7 @@ export const config = {
},
{
match: { hw: "b200", variant: "default", quant: "bf16", strategy: "balanced", nodes: "multi-2" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -887,9 +847,7 @@ export const config = {
},
{
match: { hw: "b200", variant: "default", quant: "bf16", strategy: "mtp", nodes: "multi-2" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
],
env: [],
flags: [
"--trust-remote-code",
"--model-path {{MODEL_NAME}}",
@@ -925,7 +883,6 @@ export const config = {
match: { hw: "b200", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -960,7 +917,6 @@ export const config = {
match: { hw: "b300", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -994,7 +950,6 @@ export const config = {
{
match: { hw: "gb200", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1029,7 +984,6 @@ export const config = {
match: { hw: "gb300", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1064,7 +1018,6 @@ export const config = {
match: { hw: "h200", variant: "lora", quant: "nvfp4", strategy: "balanced", nodes: "single" },
verified: true,
env: [
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
],
@@ -1102,7 +1055,6 @@ export const config = {
"NCCL_MNNVL_ENABLE=1",
"NCCL_NVLS_ENABLE=1",
"NCCL_CUMEM_ENABLE=1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE=1",
"SGLANG_EXPERIMENTAL_LORA_OPTI=1",
"SGLANG_OPT_LORA_OVERLAP_MAIN_ALLOC=1",
"SGLANG_OPT_USE_JIT_KERNEL_MOE_ALIGN=1",
@@ -11,7 +11,6 @@ from sglang.srt.arg_groups.model_override_base import (
is_attention_backend_not_set,
resolving_view,
)
from sglang.srt.environ import envs
from sglang.srt.runtime_context import get_platform
logger = logging.getLogger(__name__)
@@ -76,5 +75,4 @@ def _inkling_overrides(server_args: Any, hf_config: Any) -> dict:
f"Use {inkling_attn_backend} as the attention backend for Inkling "
"(requires fa4 or triton)."
)
envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.set(True)
return overrides
+1 -6
View File
@@ -671,10 +671,6 @@ class Envs:
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS = EnvBool(True)
# Decode batches between SWA out-of-window evictions.
SGLANG_SWA_EVICTION_INTERVAL = EnvInt(128)
# Deprecated: the unified radix tree is the default tree cache now, so the
# registry no longer reads this. Kept because a few model/arch call sites
# still assert on it; do not use in new code.
SGLANG_ENABLE_UNIFIED_RADIX_TREE = EnvBool(False)
# Registered TreeCore backend serving the unified radix cache.
SGLANG_UNIFIED_RADIX_TREE_CORE_BACKEND = EnvStr("python")
SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW = EnvBool(False)
@@ -1975,8 +1971,7 @@ _DEPRECATED_ENVS: Dict[str, _DeprecatedEnv] = {
"To override the threshold, use '--min-free-slots-delay'."
),
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": _DeprecatedEnv(
note="The unified radix tree is the default tree cache now; unset this "
"env. The field is still defined for legacy call sites."
note="The unified radix tree is the default tree cache now; unset this env."
),
}
@@ -38,13 +38,14 @@ def walk_radix_cache_for_canary(
f"walk_radix_cache_for_canary does not support {cache_type.__name__}"
)
# RadixCache has no SWA tier, so every node it holds is SWA-resident and
# swa_resident_only is a no-op here.
slot_buf: list[int] = []
position_buf: list[int] = []
prev_slot_buf: list[int] = []
_walk_radix_subtree(
node=radix_cache.root_node,
radix_cache=radix_cache,
depth=0,
parent_last_slot=-1,
slot_buf=slot_buf,
@@ -52,7 +53,6 @@ def walk_radix_cache_for_canary(
prev_slot_buf=prev_slot_buf,
is_root=True,
unlocked_only=unlocked_only,
swa_resident_only=swa_resident_only,
)
slot_tensor = torch.tensor(slot_buf, dtype=torch.int64)
@@ -68,7 +68,6 @@ def walk_radix_cache_for_canary(
def _walk_radix_subtree(
*,
node: TreeNode,
radix_cache: BasePrefixCache,
depth: int,
parent_last_slot: int,
slot_buf: list[int],
@@ -76,21 +75,10 @@ def _walk_radix_subtree(
prev_slot_buf: list[int],
is_root: bool,
unlocked_only: bool,
swa_resident_only: bool,
) -> None:
node_slots = _node_slots_for_canary(node=node)
if unlocked_only:
emit_slots = not is_root and _node_is_unlocked_for_canary(
node=node, radix_cache=radix_cache
)
else:
emit_slots = not is_root
if swa_resident_only:
emit_slots = emit_slots and _node_is_swa_resident_for_canary(
node=node,
radix_cache=radix_cache,
)
emit_slots = not is_root and (not unlocked_only or node.lock_ref == 0)
chain_last_slot = parent_last_slot
for j, slot in enumerate(node_slots):
@@ -105,7 +93,6 @@ def _walk_radix_subtree(
for child in node.children.values():
_walk_radix_subtree(
node=child,
radix_cache=radix_cache,
depth=child_depth,
parent_last_slot=chain_last_slot,
slot_buf=slot_buf,
@@ -113,7 +100,6 @@ def _walk_radix_subtree(
prev_slot_buf=prev_slot_buf,
is_root=False,
unlocked_only=unlocked_only,
swa_resident_only=swa_resident_only,
)
@@ -122,25 +108,3 @@ def _node_slots_for_canary(*, node: TreeNode) -> list[int]:
if isinstance(value, torch.Tensor):
return [int(s) for s in value.tolist()]
return []
def _node_is_unlocked_for_canary(
*,
node: TreeNode,
radix_cache: BasePrefixCache,
) -> bool:
if type(radix_cache) is RadixCache:
return node.lock_ref == 0
raise NotImplementedError(
f"walk_radix_cache_for_canary does not support {type(radix_cache).__name__}"
)
def _node_is_swa_resident_for_canary(
*,
node: TreeNode,
radix_cache: BasePrefixCache,
) -> bool:
# RadixCache has no SWA tier, so every node it holds is resident.
return True
+1 -1
View File
@@ -65,7 +65,7 @@ def free_swa_out_of_window_slots(
if not req.kv.holds_kv:
return
# For swa radix cache, we need to evict the tokens that are not in the tree cache and also not in the sliding window
# For SWA-capable tree caches, we need to evict the tokens that are not in the tree cache and also not in the sliding window
assert req.kv.cache_protected_len % page_size == 0, (
"cache_protected_len must be page aligned"
)
@@ -36,7 +36,6 @@ from sglang.srt.configs.hybrid_arch import (
)
from sglang.srt.configs.model_config import ModelImpl, is_deepseek_dsa
from sglang.srt.environ import envs
from sglang.srt.hardware_backend.mlx.runtime import use_mlx
from sglang.srt.managers.mm_schedule import init_mm_embedding_cache
from sglang.srt.mem_cache.base_swa_memory_pool import BaseSWAKVPool
from sglang.srt.mem_cache.cache_init_params import CacheInitParams
@@ -314,12 +313,6 @@ def build_kv_cache(
and get_disagg().disaggregation_mode == "decode"
):
if is_hybrid_swa:
if not (envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.get() or use_mlx()):
raise ValueError(
"--disaggregation-decode-enable-radix-cache with sliding "
"window attention (SWA) models requires the unified radix "
"tree (set SGLANG_ENABLE_UNIFIED_RADIX_TREE=1)."
)
if enable_hierarchical_cache:
raise ValueError(
"--disaggregation-decode-enable-radix-cache with sliding "
@@ -367,18 +367,6 @@ class UnifiedLRUList:
return None
return x
def get_prev_leaf_no_lock(self, node: UnifiedTreeNode, check_id: bool = True):
if check_id:
assert node.id in self.cache
pt = self._pt
ct = self.component_type
x = node.lru_prev[pt]
while x.component_data[ct].lock_ref > 0 or len(x.children) > 0:
x = x.lru_prev[pt]
if x == self.head:
return None
return x
def get_prev_no_host_lock(self, node: UnifiedTreeNode, check_id: bool = True):
"""Host-LRU walker: skip nodes whose component host_lock_ref > 0."""
if check_id:
@@ -395,9 +383,6 @@ class UnifiedLRUList:
def get_lru_no_lock(self):
return self.get_prev_no_lock(self.tail, check_id=False)
def get_leaf_lru_no_lock(self):
return self.get_prev_leaf_no_lock(self.tail, check_id=False)
def get_lru_no_host_lock(self):
return self.get_prev_no_host_lock(self.tail, check_id=False)
-1
View File
@@ -1014,7 +1014,6 @@ class InklingForConditionalGeneration(nn.Module):
self.config = config
self.text_config = config.text_config
assert envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.get()
if get_disagg().disaggregation_mode != "decode":
assert not get_memory().disable_radix_cache
assert not get_schedule().disable_hybrid_swa_memory
-1
View File
@@ -668,7 +668,6 @@ def unified_radix_tree_server_env(
**os.environ,
**extra_env,
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
"SGLANG_UNIFIED_RADIX_TREE_CORE_BACKEND": tree_core_backend,
}
@@ -44,9 +44,6 @@ class TestDisaggregationDecodeRadixCacheSWANixl(
# so keep the original 0.45 absolute floor and rely on the two-pass
# non-regression check below to catch decode-cache corruption.
gsm8k_min_score = 0.45
# SWA + decode-side radix cache is gated to the unified radix tree.
extra_prefill_env = {"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"}
extra_decode_env = {"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"}
extra_prefill_args = SWA_SERVER_ARGS
extra_decode_args = [
"--disaggregation-decode-enable-radix-cache",
@@ -28,10 +28,6 @@ MODEL = os.environ.get(
"INKLING_SMALL_TEST_MODEL_PATH", "thinkingmachines/Inkling-Small-NVFP4"
)
# The unified radix tree is what merges the three components into one tree, so
# it is a precondition rather than a tuning knob here.
ENV = {"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"}
# Shared with the single-server Inkling recipe; TP is per role.
COMMON_ARGS = [
"--tp",
@@ -104,7 +100,6 @@ class TestDisaggregationInklingMXFP8(PDDisaggregationServerBase, GSM8KMixin):
cls.prefill_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=prefill_args,
env={**os.environ, **ENV},
)
@classmethod
@@ -125,7 +120,6 @@ class TestDisaggregationInklingMXFP8(PDDisaggregationServerBase, GSM8KMixin):
cls.decode_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=decode_args,
env={**os.environ, **ENV},
)
@@ -37,9 +37,6 @@ register_cuda_ci(est_time=900, stage="extra-a", runner_config="2-gpu-large")
_MODEL_PATH = os.environ.get("INKLING_TEST_MODEL_PATH", "thinkingmachines/Inkling")
_MODEL_REVISION = os.environ.get("INKLING_TEST_MODEL_REVISION", "test")
# The unified radix tree is what merges the three components into one tree.
SERVER_ENV = {"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"}
UNIFIED_TRI_ARGS = [
"--skip-tokenizer-init",
"--random-seed",
@@ -72,8 +69,6 @@ class TestUnifiedMemoryDisaggregationTriPool(
"""1 prefill + 1 decode, both unified, vs a non-PD unified reference."""
model = _MODEL_PATH
extra_prefill_env = SERVER_ENV
extra_decode_env = SERVER_ENV
prefill_tp_size = 1
decode_tp_size = 1
decode_base_gpu_id = 1
@@ -6,7 +6,6 @@ same unified-memory configuration to keep attention reduction order comparable.
Covers GDN, SWA, tri-pool, and MLA layouts.
"""
import os
import time
import unittest
@@ -67,7 +66,6 @@ class UnifiedMemoryHiCacheBase(CustomTestCase):
model: str = ""
extra_args: list = []
server_env: dict = {}
@classmethod
def setUpClass(cls):
@@ -76,7 +74,6 @@ class UnifiedMemoryHiCacheBase(CustomTestCase):
base_args = _COMMON_ARGS + cls.extra_args
cls.hicache_url = "http://127.0.0.1:8157"
cls.reference_url = "http://127.0.0.1:8158"
env = {**os.environ, **cls.server_env} if cls.server_env else None
hicache_args = ["--enable-hierarchical-cache"]
if "--hicache-size" not in base_args:
hicache_args += ["--hicache-ratio", "4"]
@@ -85,7 +82,6 @@ class UnifiedMemoryHiCacheBase(CustomTestCase):
cls.hicache_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=base_args + hicache_args,
env=env,
)
cls.addClassCleanup(kill_process_tree, cls.process_hicache.pid)
cls.process_reference = popen_launch_server(
@@ -93,7 +89,6 @@ class UnifiedMemoryHiCacheBase(CustomTestCase):
cls.reference_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=base_args + ["--base-gpu-id", "1"],
env=env,
)
cls.addClassCleanup(kill_process_tree, cls.process_reference.pid)
@@ -188,7 +183,6 @@ class TestUnifiedMemoryHiCacheTriPool(UnifiedMemoryHiCacheBase):
# The test revision is the reduced checkpoint used by Inkling CI.
model = "thinkingmachines/Inkling"
server_env = {"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"}
extra_args = _SMALL_POOL + [
"--revision",
"test",
@@ -86,7 +86,6 @@ class TestInklingServer(CustomTestCase):
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
@classmethod
@@ -112,7 +112,6 @@ class TestInklingSmallNvfp4(CustomTestCase):
"--mem-fraction-static",
"0.85",
],
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
@classmethod
@@ -202,7 +201,6 @@ class TestInklingSmallNvfp4DsparkDeterministic(CustomTestCase):
"--speculative-draft-attention-backend",
"fa4",
],
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
@classmethod
@@ -316,7 +314,6 @@ class TestInklingSmallNvfp4HiCacheDeterministic(CustomTestCase):
"4",
"--enable-deterministic-inference",
],
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18, trust_remote_code=True)
@@ -141,7 +141,6 @@ class TestInklingUnifiedTriPool(CustomTestCase):
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=cls.server_args(),
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
@classmethod
@@ -235,7 +234,6 @@ class TestInklingUnifiedVsStaticParity(CustomTestCase):
DEFAULT_URL_FOR_TEST,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
try:
out = []
@@ -1,6 +1,5 @@
import unittest
from sglang.srt.environ import envs
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
@@ -50,11 +49,6 @@ class TestMiMoV2(GSM8KMixin, MMMUServerBase):
server_api_key = None
other_args = MIMO_V2_MTP_OTHER_ARGS
@classmethod
def setUpClass(cls):
with envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.override(True):
super().setUpClass()
if __name__ == "__main__":
unittest.main()
@@ -1,6 +1,5 @@
import unittest
from sglang.srt.environ import envs
from sglang.srt.utils import is_blackwell
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
@@ -54,11 +53,6 @@ class TestMiMoV2Flash(GSM8KMixin, SpecDecodingMixin, DefaultServerBase):
bs_1_speed_thres = 170
accept_length_thres = 3.2
@classmethod
def setUpClass(cls):
with envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.override(True):
super().setUpClass()
if __name__ == "__main__":
unittest.main()
@@ -1,6 +1,5 @@
import unittest
from sglang.srt.environ import envs
from sglang.test.accuracy_test_runner import AccuracyTestParams
from sglang.test.cache_consistency_jitter import get_jitter_engine, run_jitter_test
from sglang.test.ci.ci_register import register_cuda_ci
@@ -58,24 +57,23 @@ class TestInklingNVFP4Nightly(unittest.TestCase):
),
]
with envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.override(1):
run_combined_tests(
models=variants,
test_name="Inkling-NVFP4",
accuracy_params=AccuracyTestParams(
dataset="gsm8k",
baseline_accuracy=GSM8K_BASELINE,
num_examples=1314,
num_threads=512,
max_tokens=16000,
temperature=1.0,
top_p=0.95,
repeat=1,
),
performance_params=PerformanceTestParams(
result_dir="performance_results_inkling_nvfp4",
),
)
run_combined_tests(
models=variants,
test_name="Inkling-NVFP4",
accuracy_params=AccuracyTestParams(
dataset="gsm8k",
baseline_accuracy=GSM8K_BASELINE,
num_examples=1314,
num_threads=512,
max_tokens=16000,
temperature=1.0,
top_p=0.95,
repeat=1,
),
performance_params=PerformanceTestParams(
result_dir="performance_results_inkling_nvfp4",
),
)
class TestInklingSmallCacheConsistencyNightly(unittest.TestCase):
@@ -85,27 +83,26 @@ class TestInklingSmallCacheConsistencyNightly(unittest.TestCase):
@unittest.skipIf(not is_blackwell_system(), "NVFP4 requires Blackwell")
def test_scored_contexts_are_bitwise_identical(self):
with envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.override(1):
with get_jitter_engine(
model_path=INKLING_SMALL_NVFP4_MODEL,
tp_size=8,
trust_remote_code=True,
quantization="modelopt_fp4",
attention_backend="fa4",
page_size=128,
fp4_gemm_runner_backend="flashinfer_trtllm",
moe_runner_backend="flashinfer_trtllm_routed",
enable_torch_symm_mem=True,
mamba_radix_cache_strategy="extra_buffer",
swa_full_tokens_ratio=0.1,
mamba_full_memory_ratio=0.1,
mem_fraction_static=0.6,
enable_deterministic_inference=True,
# The harness pins a tight pool; this checkpoint needs room for
# the 2048-4096 token prefixes the default workload draws.
max_total_tokens=131_072,
) as engine:
run_jitter_test(engine)
with get_jitter_engine(
model_path=INKLING_SMALL_NVFP4_MODEL,
tp_size=8,
trust_remote_code=True,
quantization="modelopt_fp4",
attention_backend="fa4",
page_size=128,
fp4_gemm_runner_backend="flashinfer_trtllm",
moe_runner_backend="flashinfer_trtllm_routed",
enable_torch_symm_mem=True,
mamba_radix_cache_strategy="extra_buffer",
swa_full_tokens_ratio=0.1,
mamba_full_memory_ratio=0.1,
mem_fraction_static=0.6,
enable_deterministic_inference=True,
# The harness pins a tight pool; this checkpoint needs room for
# the 2048-4096 token prefixes the default workload draws.
max_total_tokens=131_072,
) as engine:
run_jitter_test(engine)
if __name__ == "__main__":
@@ -32,9 +32,6 @@ register_cuda_ci(est_time=308, stage="base-b", runner_config="2-gpu-large")
class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
# Extra server env; subclasses use it to select a tree_cache implementation.
extra_env: dict = {}
@classmethod
def setUpClass(cls):
cls.temp_dir = tempfile.mkdtemp()
@@ -64,7 +61,6 @@ class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
# Make runs less flaky for CI/dev.
"SGLANG_ENABLE_DETERMINISTIC_INFERENCE": "1",
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
**cls.extra_env,
}
@classmethod
@@ -372,18 +368,5 @@ class TestHiCacheStorageRuntimeAttachDetach(CustomTestCase):
time.sleep(2)
class TestUnifiedRadixCacheStorageRuntimeAttachDetach(
TestHiCacheStorageRuntimeAttachDetach
):
"""Same runtime attach/detach lifecycle, backed by UnifiedRadixCache."""
extra_env = {"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"}
def test_runtime_attach_detach(self):
# Admin-key gating (phase A of the base test) lives in the HTTP layer and is
# independent of the tree cache implementation, so only the lifecycle is run.
self._check_attach_detach_lifecycle()
if __name__ == "__main__":
unittest.main(verbosity=2)
@@ -126,7 +126,6 @@ class TestHiCacheStorageUMBPBackend(CustomTestCase):
{
"SGLANG_ENABLE_DETERMINISTIC_INFERENCE": "1",
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
"SGLANG_DSV4_FP4_EXPERTS": "0",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_USE_ROCM700A": "0",
@@ -21,7 +21,6 @@ Usage:
python3 -m unittest test_int8_mamba_checkpoint_e2e
"""
import time
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
@@ -30,21 +29,19 @@ from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
from sglang.test.server_fixtures.default_fixture import (
DefaultServerBase,
openai_api_env,
)
from sglang.test.test_utils import (
DEFAULT_HYBRID_MAMBA_MODEL_NAME_FOR_TEST,
popen_launch_server,
terminate_and_kill_process_tree,
)
register_cuda_ci(est_time=344, stage="extra-b", runner_config="4-gpu-h100")
register_cuda_ci(est_time=172, stage="extra-b", runner_config="4-gpu-h100")
class TestInt8MambaCheckpointE2E(KLDivergenceMixin, DefaultServerBase):
"""int8 mamba checkpoint pool on Qwen3-Next-80B-A3B (GDN-hybrid)."""
model = DEFAULT_HYBRID_MAMBA_MODEL_NAME_FOR_TEST
server_env = {"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1"}
# Cache-hit KL: int8 is a lossy codec, so its cache-hit divergence is
# inherently larger than the bf16/fp8 reuse the other KL tests bound (~0.005),
@@ -97,30 +94,5 @@ class TestInt8MambaCheckpointE2E(KLDivergenceMixin, DefaultServerBase):
self.assertGreaterEqual(metrics["accuracy"], self.gsm8k_threshold)
class TestUnifiedRadixTreeInt8MambaCheckpointE2E(TestInt8MambaCheckpointE2E):
"""Run the same int8 mamba checkpoint checks with UnifiedRadixTree forced on."""
@classmethod
def setUpClass(cls):
assert cls.model is not None, "Please set cls.model in subclass"
with openai_api_env(cls.api_key):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=cls.timeout,
other_args=cls.other_args,
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
@classmethod
def tearDownClass(cls):
terminate_and_kill_process_tree(cls.process, wait_timeout=60)
time.sleep(2)
if __name__ == "__main__":
unittest.main()
@@ -73,7 +73,6 @@ class TestDeepSeekV4FlashUnifiedCacheLinkerKL(
env={
**cls.mooncake.server_env(),
"SGLANG_DSV4_FP4_EXPERTS": "0",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18)
@@ -101,7 +101,6 @@ class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18)
@@ -154,7 +153,6 @@ class TestUnifiedQwen3HiCachePPL3(AccuracyTwoPassMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.hicache_dir,
},
)
@@ -65,7 +65,6 @@ class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18)
@@ -92,7 +92,6 @@ class TestUnifiedKimiLinearDcpHiCache(UnifiedRadixTreeTestMixin, CustomTestCase)
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18, trust_remote_code=True)
@@ -71,7 +71,6 @@ class TestGLM5UnifiedRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
env={
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.hicache_dir,
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
@@ -65,7 +65,6 @@ class TestUnifiedMambaRadixCache(UnifiedRadixTreeTestMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18)
@@ -127,7 +126,6 @@ class TestUnifiedMambaHiCache(UnifiedRadixTreeTestMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18)
@@ -197,7 +195,6 @@ class TestUnifiedMambaHiCacheL3(AccuracyTwoPassMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.hicache_dir,
},
)
@@ -91,7 +91,6 @@ class TestUnifiedMiMoHiCacheLoadBackKL(CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
"SGLANG_USE_CUDA_IPC_TRANSPORT": "1",
},
)
@@ -39,7 +39,6 @@ class TestUnifiedSWARadixCache(UnifiedRadixTreeTestMixin, CustomTestCase):
],
env={
"SGLANG_ENABLE_RANK_CONSENSUS_CHECKER": "1",
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
},
)
cls.input_ids = get_input_ids(cls.model, num_samples=18)
@@ -29,10 +29,6 @@ register_amd_ci(est_time=9, suite="stage-b-test-1-gpu-small-amd")
register_xpu_ci(est_time=20, suite="stage-b-test-1-gpu-xpu")
def _event_hashes(events):
return [block_hash for event in events for block_hash in event.block_hashes]
class TestMamba(unittest.TestCase):
@classmethod
def setUpClass(cls):
@@ -364,7 +364,7 @@ class TestDefaultRadixCacheFactory(CustomTestCase):
)
self.assertIs(result, cache)
def test_swa_radix_cache_when_hybrid_swa(self):
def test_unified_radix_cache_when_hybrid_swa(self):
ctx = _make_ctx(self, is_hybrid_swa=True)
# SWA hybrid models now default to the unified radix tree.
fake_components = MagicMock()
@@ -390,7 +390,7 @@ class TestDefaultRadixCacheFactory(CustomTestCase):
PureSWA.assert_called_once_with(params=ctx.params)
self.assertIs(result, PureSWA.return_value)
def test_mamba_radix_cache_when_hybrid_ssm(self):
def test_unified_radix_cache_when_hybrid_ssm(self):
ctx = _make_ctx(self, is_hybrid_ssm=True)
# Mamba hybrid models now default to the unified radix tree.
fake_components = MagicMock()
@@ -27,22 +27,12 @@ register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-large")
register_amd_ci(est_time=10, suite="stage-b-test-1-gpu-small-amd")
class _DummyReq:
def __init__(self):
self._kv_committed_len = 0
self.swa_prefix_lock_released = False
self.kv = SimpleNamespace(swa_evicted_seqlen=0, cache_protected_len=0)
def _build_swa_tree(
is_eagle: bool,
def _build_swa_allocator(
page_size: int = 1,
req_size: int = 8,
max_context_len: int = 64,
kv_size: int = 64,
kv_size_swa: int = 32,
sliding_window_size: int = 4,
enable_kv_cache_events: bool = False,
swa_req_ring_size: int | None = None,
):
head_num = 8
@@ -149,12 +139,10 @@ class TestSWA(unittest.TestCase):
def test_swa_memory_pool_paged_free_clears_full_page_mapping(self):
page_size = 4
allocator, _ = _build_swa_tree(
is_eagle=False,
allocator, _ = _build_swa_allocator(
page_size=page_size,
kv_size=16,
kv_size_swa=16,
sliding_window_size=page_size,
)
full_indices = _swa_alloc(allocator, page_size)
@@ -176,7 +164,7 @@ class TestSWA(unittest.TestCase):
"""Clearing the full-to-SWA mapping must not block the stream; writing a
host-resident scalar into it does.
"""
allocator, _ = _build_swa_tree(is_eagle=False)
allocator, _ = _build_swa_allocator()
full_indices = _swa_alloc(allocator, 4)
mapping = allocator.full_to_swa_index_mapping
@@ -202,8 +190,7 @@ class TestSWA(unittest.TestCase):
self._free_swa_group_owns_deferred_indices(page_size)
def _free_swa_group_owns_deferred_indices(self, page_size):
allocator, _ = _build_swa_tree(
is_eagle=False,
allocator, _ = _build_swa_allocator(
page_size=page_size,
kv_size=32 * page_size,
kv_size_swa=32 * page_size,
@@ -239,8 +226,7 @@ class TestSWA(unittest.TestCase):
)
def test_free_swa_group_owns_mapping_at_enqueue_time(self):
allocator, _ = _build_swa_tree(
is_eagle=False,
allocator, _ = _build_swa_allocator(
kv_size=8,
kv_size_swa=8,
)
@@ -271,8 +257,7 @@ class TestSWA(unittest.TestCase):
)
def _build_two_mapped_slots(self, page_size=1):
allocator, _ = _build_swa_tree(
is_eagle=False,
allocator, _ = _build_swa_allocator(
page_size=page_size,
kv_size=8 * page_size,
kv_size_swa=8 * page_size,
@@ -372,7 +357,7 @@ class TestFreeFullPartition(CustomTestCase):
"""`free_full` releases only the full side of a hybrid SWA allocator."""
def setUp(self):
self.allocator, _ = _build_swa_tree(is_eagle=False)
self.allocator, _ = _build_swa_allocator()
self.full_baseline = self.allocator.full_available_size()
self.swa_baseline = self.allocator.swa_available_size()
@@ -420,7 +405,7 @@ class TestFreeKvRow(CustomTestCase):
whole, the SWA side only from the floor up."""
def setUp(self):
self.allocator, _ = _build_swa_tree(is_eagle=False)
self.allocator, _ = _build_swa_allocator()
self.full_baseline = self.allocator.full_available_size()
self.swa_baseline = self.allocator.swa_available_size()
@@ -457,7 +442,7 @@ class TestFreeKvRow(CustomTestCase):
self.assertEqual(self._sizes(), (self.full_baseline, self.swa_baseline))
def test_below_floor_pieces_go_back_through_the_full_side(self):
allocator, _ = _build_swa_tree(is_eagle=False, page_size=4)
allocator, _ = _build_swa_allocator(page_size=4)
indices = _swa_alloc(allocator, 8)
allocator.free_swa(indices)
after_alloc = allocator.full_available_size()
@@ -477,7 +462,7 @@ class TestFreeKvRow(CustomTestCase):
self.assertEqual(allocator.full_available_size(), after_alloc + 8)
def test_grouped_full_side_frees_defer_and_skip_the_unique_path(self):
allocator, _ = _build_swa_tree(is_eagle=False, page_size=4)
allocator, _ = _build_swa_allocator(page_size=4)
indices = _swa_alloc(allocator, 12)
allocator.free_swa(indices[:8])
after_alloc = allocator.full_available_size()
@@ -540,7 +525,7 @@ class TestSWAPeerMappedContract(CustomTestCase):
return bool(assert_async.call_args.args[0])
def test_segment_free_flags_a_page_whose_peer_is_already_gone(self):
allocator, _ = _build_swa_tree(is_eagle=False, page_size=4)
allocator, _ = _build_swa_allocator(page_size=4)
live = _swa_alloc(allocator, 8)
stale = _swa_alloc(allocator, 8)
allocator.clear_full_to_swa_mapping(stale)
@@ -553,7 +538,7 @@ class TestSWAPeerMappedContract(CustomTestCase):
"""page_size > 1: page reps by stride replace the page expansion's
filter and the inner allocator's torch.unique, in and out of a group."""
ps = 4
allocator, _ = _build_swa_tree(is_eagle=False, page_size=ps)
allocator, _ = _build_swa_allocator(page_size=ps)
def grouped(indices):
allocator.free_group_begin()
@@ -581,7 +566,7 @@ class TestSWAPeerMappedContract(CustomTestCase):
self.assertIsNone(_sync_error(lambda: grouped(second[: 2 * ps - 1])))
def test_free_swa_flags_a_slot_whose_peer_is_already_gone(self):
allocator, _ = _build_swa_tree(is_eagle=False)
allocator, _ = _build_swa_allocator()
live = _swa_alloc(allocator, 4)
stale = _swa_alloc(allocator, 4)
# Whoever released the peer left the mapping reading as the padding slot.
@@ -594,7 +579,7 @@ class TestSWAPeerMappedContract(CustomTestCase):
def test_free_swa_does_not_synchronize(self):
"""The filter's output shape was data-dependent, so it read a count back
to the host; the gather that replaced it has a fixed shape."""
allocator, _ = _build_swa_tree(is_eagle=False)
allocator, _ = _build_swa_allocator()
mapping = allocator.full_to_swa_index_mapping
# Warm up outside the window: a first-time cudaMalloc can synchronize on
@@ -618,8 +603,7 @@ class TestSWAReqRingFree(CustomTestCase):
def _allocated_ring(self):
ps = self.PS
allocator, req_pool = _build_swa_tree(
is_eagle=False,
allocator, req_pool = _build_swa_allocator(
page_size=ps,
req_size=2,
max_context_len=4 * ps,
@@ -738,7 +722,7 @@ class TestSWAPageRepsFree(CustomTestCase):
PS = 4
def _allocator(self):
allocator, _ = _build_swa_tree(is_eagle=False, page_size=self.PS)
allocator, _ = _build_swa_allocator(page_size=self.PS)
return allocator
def _sizes(self, allocator):
@@ -748,8 +732,7 @@ class TestSWAPageRepsFree(CustomTestCase):
def test_free_swa_segment_npu_uses_reference_path(self):
for page_size in (1, 4):
with self.subTest(page_size=page_size):
allocator, _ = _build_swa_tree(
is_eagle=False,
allocator, _ = _build_swa_allocator(
page_size=page_size,
kv_size=8 * page_size,
kv_size_swa=8 * page_size,
@@ -800,8 +783,7 @@ class TestSWAPageRepsFree(CustomTestCase):
with self.subTest(name=name):
num_allocated_tokens = max(2 * page_size, num_tokens)
kv_size = max(8 * page_size, num_allocated_tokens)
allocator, _ = _build_swa_tree(
is_eagle=False,
allocator, _ = _build_swa_allocator(
page_size=page_size,
kv_size=kv_size,
kv_size_swa=kv_size,