Viewing enriched Twitter post
Why do we store the SSM state at all? More and more models are hybrids (Nemotron-3, Qwen3.5), so SSM decode speed matters. We only write it back every step so the next step can read it. ReplaySSM caches the recent inputs instead and rebuilds the state on the fly. Same outputs, half the memory traffic → ~2x on spec decode at large batch sizes, which barely even helped SSMs before → up to 1.43x standard decode on large hybrids (up to Nemotron-Ultra-550B) Work with @tri_dao Blog + Code👇
{
"media": [
{
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2066517132733509756/media_0.jpg",
"media_url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2066517132733509756/media_0.jpg",
"type": "photo",
"filename": "media_0.jpg"
}
],
"processed_at": "2026-06-15T13:50:31.657565",
"pipeline_version": "2.0"
}{
"type": "tweet",
"id": "2066517132733509756",
"url": "https://x.com/zwljohnny/status/2066517132733509756",
"twitterUrl": "https://twitter.com/zwljohnny/status/2066517132733509756",
"text": "Why do we store the SSM state at all?\nMore and more models are hybrids (Nemotron-3, Qwen3.5), so SSM decode speed matters.\n\nWe only write it back every step so the next step can read it.\nReplaySSM caches the recent inputs instead and rebuilds the state on the fly.\n\nSame outputs, half the memory traffic\n→ ~2x on spec decode at large batch sizes, which barely even helped SSMs before\n→ up to 1.43x standard decode on large hybrids (up to Nemotron-Ultra-550B)\n\nWork with @tri_dao\n\nBlog + Code👇",
"source": "Twitter for iPhone",
"retweetCount": 0,
"replyCount": 1,
"likeCount": 1,
"quoteCount": 1,
"viewCount": 11,
"createdAt": "Mon Jun 15 13:44:13 +0000 2026",
"lang": "en",
"bookmarkCount": 0,
"isReply": false,
"inReplyToId": null,
"conversationId": "2066517132733509756",
"displayTextRange": [
0,
279
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "zwljohnny",
"url": "https://x.com/zwljohnny",
"twitterUrl": "https://twitter.com/zwljohnny",
"id": "2019902551391350785",
"name": "Ze-Wei (Johnny) Liou",
"isVerified": false,
"isBlueVerified": false,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/2052115348443275265/pxNseCff_normal.jpg",
"coverPicture": "",
"description": "",
"location": "Princeton, NJ",
"followers": 4,
"following": 32,
"status": "",
"canDm": false,
"canMediaTag": true,
"createdAt": "Fri Feb 06 22:34:53 +0000 2026",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 2,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 7,
"statusesCount": 7,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [],
"profile_bio": {
"description": "First year PhD @Princeton working on ML Systems",
"entities": {
"description": {
"user_mentions": [
{
"id_str": "",
"indices": [
15,
25
],
"name": "",
"screen_name": "Princeton"
}
]
},
"url": {
"urls": [
{
"display_url": "johnny-liou.github.io",
"expanded_url": "http://johnny-liou.github.io",
"indices": [
0,
23
],
"url": "https://t.co/cEt3GicemF"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"allow_download_status": {
"allow_download": true
},
"display_url": "pic.twitter.com/wRoeDT6jx3",
"expanded_url": "https://twitter.com/zwljohnny/status/2066517132733509756/photo/1",
"ext_master_playlist_only": [],
"ext_media_availability": {
"status": "Available"
},
"ext_playlists": [],
"features": {
"large": {
"faces": []
},
"orig": {
"faces": []
}
},
"id_str": "2066517129709387776",
"indices": [
280,
303
],
"media_key": "3_2066517129709387776",
"media_results": {
"id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARytvmFuloAACgACHK2+YiLW8HwAAA==",
"result": {
"__typename": "ApiMedia",
"id": "QXBpTWVkaWE6DAABCgABHK2+YW6WgAAKAAIcrb5iItbwfAAA",
"media_key": "3_2066517129709387776"
}
},
"media_url_https": "https://pbs.twimg.com/media/HK2-YW6WgAA5uvr.jpg",
"original_info": {
"focus_rects": [
{
"h": 520,
"w": 929,
"x": 271,
"y": 0
},
{
"h": 520,
"w": 520,
"x": 667,
"y": 0
},
{
"h": 520,
"w": 456,
"x": 699,
"y": 0
},
{
"h": 520,
"w": 260,
"x": 797,
"y": 0
},
{
"h": 520,
"w": 1200,
"x": 0,
"y": 0
}
],
"height": 520,
"width": 1200
},
"sizes": {
"large": {
"h": 520,
"w": 1200
}
},
"type": "photo",
"url": "https://t.co/wRoeDT6jx3"
}
]
},
"card": null,
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"urls": [],
"user_mentions": [
{
"id_str": "568879807",
"indices": [
470,
478
],
"name": "Tri Dao",
"screen_name": "tri_dao"
}
]
},
"quoted_tweet": null,
"retweeted_tweet": null,
"isLimitedReply": false,
"communityInfo": null,
"article": null
}