[Spec] Redo: split init_backends; account draft weights in --mem-fraction-static (#28855)

Co-authored-by: cctry <cctry@fb.com>

test/registered/8-gpu-models/test_deepseek_v32.py

@@ -53,7 +53,7 @@ class TestDeepseekV32(unittest.TestCase):
             "--speculative-num-steps=3",
             "--speculative-eagle-topk=1",
             "--speculative-num-draft-tokens=4",
-            "--mem-frac=0.7",
+            "--mem-frac=0.85",
         ]
         variants = [
             # Variant: "dp" - Standard TP=8 + DP=8 with dp-attention

test/registered/cp/test_deepseek_v32_cp_single_node.py

@@ -43,7 +43,7 @@ class TestDeepseekV32CPInSeqSplit(CustomTestCase):
             "--speculative-num-draft-tokens",
             "4",
             "--mem-frac",
-            "0.7",
+            "0.85",
             "--cuda-graph-max-bs",
             "32",
             "--max-running-requests",
@@ -109,7 +109,7 @@ class TestDeepseekV32CPRoundRobinSplit(CustomTestCase):
             "--speculative-num-draft-tokens",
             "4",
             "--mem-frac",
-            "0.7",
+            "0.85",
             "--cuda-graph-max-bs",
             "32",
             "--max-running-requests",

test/registered/models_e2e/test_deepseek_v3_mtp.py

@@ -31,7 +31,7 @@ _OTHER_ARGS = [
     '{"enable_multithread_load": true, "num_threads": 64}',
 ]
 if not is_in_amd_ci():
-    _OTHER_ARGS += ["--mem-frac", "0.7"]
+    _OTHER_ARGS += ["--mem-frac", "0.85"]
 
 
 class TestDeepseekV3MTP(GSM8KMixin, DefaultServerBase):

test/registered/models_e2e/test_dsa_dsv32_dp_mtp.py

@@ -19,7 +19,7 @@ class TestDeepseekV32DPMTP(
     DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin
 ):
     model = "deepseek-ai/DeepSeek-V3.2"
-    mem_fraction_static = 0.7
+    mem_fraction_static = 0.85
     enable_dp_attention = True
     bs_1_speed_thres = 90
 

test/registered/models_e2e/test_dsa_dsv32_tp_mtp.py

@@ -19,7 +19,7 @@ class TestDeepseekV32TPMTP(
     DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin
 ):
     model = "deepseek-ai/DeepSeek-V3.2"
-    mem_fraction_static = 0.7
+    mem_fraction_static = 0.85
     bs_1_speed_thres = 180
 
 

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dsv4.py

@@ -200,7 +200,7 @@ class TestUnifiedDeepSeekV4FlashEagleHiCacheL3(AccuracyTwoPassMixin, CustomTestC
                 "--chunked-prefill-size",
                 "8192",
                 "--mem-fraction-static",
-                "0.9",
+                "0.95",
                 "--disable-shared-experts-fusion",
                 "--enable-hierarchical-cache",
                 "--hicache-ratio",