[Docs] quick fix delete --enable-dp-attention in sgl-jax (#24052)
This commit is contained in:
@@ -230,6 +230,12 @@ export const MiMoV25Deployment = () => {
|
|||||||
// Recipe sources:
|
// Recipe sources:
|
||||||
// v7x: tp=ep=32, dp=4, omits --attention-backend, mem-frac 0.95, swa 0.25
|
// v7x: tp=ep=32, dp=4, omits --attention-backend, mem-frac 0.95, swa 0.25
|
||||||
// v6e: tp=ep=64, dp=8, --attention-backend fa, mem-frac 0.92, swa 0.15
|
// v6e: tp=ep=64, dp=8, --attention-backend fa, mem-frac 0.92, swa 0.15
|
||||||
|
//
|
||||||
|
// sgl-jax conventions:
|
||||||
|
// - `--tp-size` is always the total JAX device count; per-DP TP is
|
||||||
|
// derived automatically as tp/dp.
|
||||||
|
// - No `--enable-dp-attention` flag — DP attention is the default
|
||||||
|
// (FFN layers auto-pick EP-split for MoE, attn-TP-split for dense).
|
||||||
const isV7x = hardware === "tpu-v7x";
|
const isV7x = hardware === "tpu-v7x";
|
||||||
const useEp = expertParallelism === "enabled";
|
const useEp = expertParallelism === "enabled";
|
||||||
const useDpAttn = dpAttention === "enabled";
|
const useDpAttn = dpAttention === "enabled";
|
||||||
@@ -239,7 +245,7 @@ export const MiMoV25Deployment = () => {
|
|||||||
flags.push(" --trust-remote-code");
|
flags.push(" --trust-remote-code");
|
||||||
flags.push(` --tp-size ${tp}`);
|
flags.push(` --tp-size ${tp}`);
|
||||||
if (useEp) flags.push(` --ep-size ${tp}`);
|
if (useEp) flags.push(` --ep-size ${tp}`);
|
||||||
if (useDpAttn) flags.push(` --dp-size ${dpSize}`, " --enable-dp-attention");
|
if (useDpAttn) flags.push(` --dp-size ${dpSize}`);
|
||||||
flags.push(" --moe-backend fused");
|
flags.push(" --moe-backend fused");
|
||||||
if (!isV7x) flags.push(" --attention-backend fa");
|
if (!isV7x) flags.push(" --attention-backend fa");
|
||||||
flags.push(" --host 0.0.0.0");
|
flags.push(" --host 0.0.0.0");
|
||||||
|
|||||||
Reference in New Issue
Block a user