minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

docs(DeepSeek-V4): mark h200|big|pd-disagg verified + recipe fixes (#23715)

Browse Source
This commit is contained in:
fzyzcjy
2026-04-25 22:49:12 +08:00
committed by GitHub
parent 046c14a3ed
commit d4c1665626
1 changed files with 31 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx
+31 -4
View File
@@ -178,8 +178,7 @@ export const DeepSeekV4Deployment = () => {
"gb300|small|max-throughput",
"h200|small|cp",
"h200|small|pd-disagg",
// h200|big|pd-disagg: pending verification (needs 4-node H200 cluster with
// shared IB fabric: 2-node prefill + 2-node decode).
"h200|big|pd-disagg",
"gb300|small|cp",
"gb300|big|cp",
"gb300|small|pd-disagg",
@@ -480,6 +479,12 @@ export const DeepSeekV4Deployment = () => {
? "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256"
: "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=1024");
}
// H200 Pro PD: tp=16 multinode + DeepEP needs the dispatch buffer cap on
// BOTH prefill + decode (matches production playground LWS for the same
// hw/model combo). Verified on 2026-04-25 (journal 2026-04-25-014).
if (hardware === "h200" && modelSize === "big") {
roleEnv.push("SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=128");
}
const envAll = [...HW_ENV, ...roleEnv, ...MNNVL_ENV, ...COMMON_ENV];
const envBlock = envAll.length ? envAll.join(" \\\n") + " \\\n" : "";
@@ -490,11 +495,33 @@ export const DeepSeekV4Deployment = () => {
flags.push(` --dp ${pdTp}`);
flags.push(" --enable-dp-attention");
if (multinode) flags.push(...multiNodeFlags(nnodes));
if (isBlackwell) flags.push(" --moe-a2a-backend deepep");
// H200 Pro PD also needs deepep: at tp=16 the FP8 block_n=128 doesn't
// divide moe intermediate_size_per_partition (3072 / 16 = 192) so MoE
// experts must be kept on a single rank rather than TP-sharded. Verified
// on 2026-04-25 (journal 2026-04-25-014, candidate cookbook Bug L).
if (isBlackwell || (hardware === "h200" && modelSize === "big")) {
flags.push(" --moe-a2a-backend deepep");
}
flags.push(` --disaggregation-mode ${mode}`);
flags.push(" --disaggregation-transfer-backend mooncake");
if (ibDevice) flags.push(` --disaggregation-ib-device ${ibDevice}`);
if (!isGB300) flags.push(` --dist-init-addr 127.0.0.1:${distPort}`);
// Same-host PD bootstrap addr; for multinode PD (h200 big tp=16 across 2
// nodes) skip this — argparse would override the multinode dist-init-addr
// already emitted by multiNodeFlags above. Verified 2026-04-25 (journal
// 2026-04-25-014). sglang falls back to its own bootstrap port (default
// 8998) which works for cross-node mooncake handshake.
if (!isGB300 && !multinode) flags.push(` --dist-init-addr 127.0.0.1:${distPort}`);
// H200 Pro PD memory-budget: cookbook defaults give available_gpu_memory
// ~17.93 GB after weights but reserve target = (1 - mem_fraction_static)
// × 138 GB = 87 GB → "Not enough memory" at memory profile. mem-frac 0.90
// and cg-max-bs 128 verified on 2026-04-25 (journal 2026-04-25-014). 128
// matches gb300|big|pd decode and gives larger decode batching headroom;
// CG capture takes ~1 hr (one-time, vs ~5 min for cg=64) but runtime
// throughput is better.
if (hardware === "h200" && modelSize === "big") {
flags.push(" --cuda-graph-max-bs 128");
flags.push(" --mem-fraction-static 0.9");
}
if (mode === "decode") {
// GB300 big PD decode is the most memory-pressured PD role: Pro 1.6T
// weights at tp=4 take ~224 GB/card on a 273 GB GB300; runtime needs