[AMD] Cap DSV4 Flash max_total_num_tokens (#30313)
Co-authored-by: YC Yen-Ching Tseng <yctseng@amd.com>
This commit is contained in:
co-authored by
YC Yen-Ching Tseng
parent
99db3b0fa5
commit
dabd4cfcfd
@@ -111,6 +111,7 @@ emit("LBPORT", rt["lb_port"])
|
|||||||
emit("MEMFRAC", rt["mem_fraction_static"])
|
emit("MEMFRAC", rt["mem_fraction_static"])
|
||||||
emit("PAGE", rt["page_size"])
|
emit("PAGE", rt["page_size"])
|
||||||
emit("MAXREQ", rt["max_running_requests"])
|
emit("MAXREQ", rt["max_running_requests"])
|
||||||
|
emit("MAXTOK", rt.get("max_total_tokens", ""))
|
||||||
emit("CHUNK", rt["chunked_prefill_size"])
|
emit("CHUNK", rt["chunked_prefill_size"])
|
||||||
# swa is DSV4-specific; emit empty when a model omits it so the flag is dropped.
|
# swa is DSV4-specific; emit empty when a model omits it so the flag is dropped.
|
||||||
emit("SWA", rt.get("swa_full_tokens_ratio", ""))
|
emit("SWA", rt.get("swa_full_tokens_ratio", ""))
|
||||||
@@ -239,6 +240,7 @@ PY
|
|||||||
EXTRA_FLAGS=""
|
EXTRA_FLAGS=""
|
||||||
(( PDP > 1 )) && EXTRA_FLAGS="$EXTRA_FLAGS --enable-dp-attention --dp-size $PDP"
|
(( PDP > 1 )) && EXTRA_FLAGS="$EXTRA_FLAGS --enable-dp-attention --dp-size $PDP"
|
||||||
(( PEP > 1 )) && EXTRA_FLAGS="$EXTRA_FLAGS --ep-size $PEP"
|
(( PEP > 1 )) && EXTRA_FLAGS="$EXTRA_FLAGS --ep-size $PEP"
|
||||||
|
[[ -n "$MAXTOK" ]] && EXTRA_FLAGS="$EXTRA_FLAGS --max-total-tokens $MAXTOK"
|
||||||
if [[ "$MTP_ENABLED" == "1" ]]; then
|
if [[ "$MTP_ENABLED" == "1" ]]; then
|
||||||
EXTRA_FLAGS="$EXTRA_FLAGS --speculative-algorithm $MTP_ALGO \
|
EXTRA_FLAGS="$EXTRA_FLAGS --speculative-algorithm $MTP_ALGO \
|
||||||
--speculative-num-steps $MTP_STEPS --speculative-eagle-topk $MTP_TOPK \
|
--speculative-num-steps $MTP_STEPS --speculative-eagle-topk $MTP_TOPK \
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -35,6 +35,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
@@ -33,6 +33,7 @@ runtime:
|
|||||||
mem_fraction_static: 0.90
|
mem_fraction_static: 0.90
|
||||||
page_size: 256
|
page_size: 256
|
||||||
max_running_requests: 256
|
max_running_requests: 256
|
||||||
|
max_total_tokens: 8551168
|
||||||
chunked_prefill_size: 8192
|
chunked_prefill_size: 8192
|
||||||
swa_full_tokens_ratio: 0.1
|
swa_full_tokens_ratio: 0.1
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user