Fix help

and correct an old value

Author: Nexesenex

KCPP-F_python_launch - help.bat

@@ -0,0 +1,3 @@
+cd /d "%~dp0"
+python koboldcpp.py --help
+pause

koboldcpp.py

@@ -5012,9 +5012,10 @@ def range_checker(arg: str):
 
     #more advanced params
     advparser = parser.add_argument_group('Advanced Commands')
-    advparser.add_argument("--ropeconfig", help="If set, uses customized RoPE scaling from configured frequency scale and frequency base (e.g. --ropeconfig 0.25 10000). Otherwise, uses NTK-Aware scaling set automatically based on context size. For NTK Rope, a rule of thumb is to double the base frequency to go 50% beyond the base context, and to triple the base frequency to double the context. Beyond, the NTK calculations are more complex and you might need to use frequency scale as well. To use only linear rope, simply set the freq-scale and ignore the freq-base",metavar=('[rope-freq-scale]', '[rope-freq-base]'), default=[0.0, 10000.0], type=float, nargs='+')
-    advparser.add_argument("--blasbatchsize", help="Sets the Logical batch size used in BLAS processing (default 128 for VRAM savings, optimal speed is 512, 256 is a great compromise). Setting it to -1 disables BLAS mode, but keeps other benefits like GPU offload. Max 4096.", type=check_range(int,-1,4096), default=128)
-    advparser.add_argument("--blasubatchsize", help="Sets the Physical batch size used in BLAS processing (default 128 for VRAM savings, optimal speed is 512, 256 is a great compromise). Setting it to 0 alignes Physical BLAS batch on logical BLAS. Max 4096.", type=check_range(int,0,4096), default=0)
+    advparser.add_argument("--ropeconfig", help="If set, uses customized RoPE scaling from configured frequency scale and frequency base (e.g. --ropeconfig 0.25 10000). Otherwise, uses NTK-Aware scaling set automatically based on context size. For NTK Rope, a rule of thumb is to double the base frequency to go 50 percents beyond the base context, and to triple the base frequency to double the context. Beyond, the NTK calculations are more complex and you might need to use frequency scale as well. To use only linear rope, simply set the freq-scale and ignore the freq-base",metavar=('[rope-freq-scale]', '[rope-freq-base]'), default=[0.0, 10000.0], type=float, nargs='+')
+    advparser.add_argument("--blasbatchsize", help="Sets the Logical batch size used in BLAS processing (default 128 for VRAM savings, optimal speed is 512, 256 is a great compromise). Setting it to -1 disables BLAS mode, but keeps other benefits like GPU offload. Steps : 1,2,4,8,16,20,24,28,32,40,48,56,64,80,96,112,128,160,192,224,256,384,512,768,1024,1536,2048,3072, max 4096.", type=check_range(int,-1,4096), default=128)
+    advparser.add_argument("--blasubatchsize", help="Sets the Physical batch size used in BLAS processing (default 128 for VRAM savings, optimal speed is 512, 256 is a great compromise). Setting it to 0 alignes Physical BLAS batch on logical BLAS. Same steps as for logical BBS.", type=check_range(int,0,4096), default=0)
+
     advparser.add_argument("--blasthreads", help="Use a different number of threads during BLAS if specified. Otherwise, has the same value as --threads",metavar=('[threads]'), type=int, default=0)
     advparser.add_argument("--lora", help="LLAMA models only, applies a lora file on top of model. Experimental.", metavar=('[lora_filename]', '[lora_base]'), nargs='+')
     advparser.add_argument("--noshift", help="If set, do not attempt to Trim and Shift the GGUF context without reprocessing everything once the max context is reached. If you disable it (or need to use Quantized KV cache (KVQ) with FlashAttention, aka. modes 1 to 20, which are incompatible with Context Shift), you can eventually use --smartcontext instead.", action='store_true')
@@ -5043,8 +5044,10 @@ def range_checker(arg: str):
     advparser.add_argument("--password", help="Enter a password required to use this instance. This key will be required for all text endpoints. Image endpoints are not secured.", default=None)
     advparser.add_argument("--ignoremissing", help="Ignores all missing non-essential files, just skipping them instead.", action='store_true')
     advparser.add_argument("--chatcompletionsadapter", help="Select an optional ChatCompletions Adapter JSON file to force custom instruct tags.", default="")
-    advparser.add_argument("--flashattention", help="Enables flash attention, which shrinks the size of the BLAS/compute cache by 50-75% (a few hundreds MB recovered in VRAM), and allows the use of the quantized KV cache.", action='store_true')
-    advparser.add_argument("--quantkv", help="Sets the KV cache data quantization (KVQ) type to save VRAM in NVidia Video Cards, 0 - F16 (16BPW) - FA or not, 1 - q8_0 - (8.5BPW) - FA, 2 - q4_0 - (4.5BPW) - FA, 3 - K F16 - V q8_0 (12.25BPW) - FA, 4 - K F16 - V q6_0 (11.25BPW) - FA, 5 - K q8_0 - V q6_0 (7.5BPW) - FA, 6 - K q8_0 - V q5_0 (7BPW) - FA, 7 - K q8_0 - V iq4_nl (6.5BPW) - FA, 8 - K q6_0 - V q6_0 (6.5BPW) - FA, 9 - K q6_0 - V q5_0 (6BPW) - FA, 10 - K q6_0 - V iq4_nl (5.5BPW) - FA, 11 - K q5_1 - V q5_0 (5.5BPW) - FA, 12 - K q5_1 - V iq4_nl (5.25BPW) - FA, 13 - K q5_0 - V iq4_nl (5BPW) - FA, 14 - K iq4_nl - V iq4_nl (4.5BPW) - FA, 15 - F16 (16BPW) - FA or not, 16 - K q8_0 - V F16 (12.25BPW) - FA or not, then slower, 17 - K q6_0 - V F16 (11.5BPW) - NO FA, slower, to be tested, 18 - K q5_1 - V F16 (11BPW) - NO FA, slower, 19 - K q5_0 - V F16 (11.75BPW) - NO FA, slower, 20 - K q4_1 - V F16 (10.5BPW) - NO FA, slower, 21 - K q4-0 - V F16 (10.25BPW) - NO FA, slower, 22 - K iq4_nl - V F16 (10.25BPW) - NO FA, slower, to be tested", metavar=('[quantization level 0/1/2/3/4/5/6/7/8/9/10/11/12/13/14/15/16/17/18/19/20/21/22]'), type=check_range(int,0,27), default=0)
+
+    advparser.add_argument("--flashattention", help="Enables flash attention, which shrinks the size of the BLAS/compute cache by 50-75 percents (a few hundreds MB recovered in VRAM), and allows the use of the quantized KV cache.", action='store_true')
+    advparser.add_argument("--quantkv", help="Sets the KV cache data quantization (KVQ) type to save VRAM in NVidia Video Cards, 0 - F16 (16BPW) - FA or not, 1 - q8_0 - (8.5BPW) - FA, 2 - q4_0 - (4.5BPW) - FA, 3 - K F16 - V q8_0 (12.25BPW) - FA, 4 - K F16 - V q6_0 (11.25BPW) - FA, 5 - K q8_0 - V q6_0 (7.5BPW) - FA, 6 - K q8_0 - V q5_0 (7BPW) - FA, 7 - K q8_0 - V iq4_nl (6.5BPW) - FA, 8 - K q6_0 - V q6_0 (6.5BPW) - FA, 9 - K q6_0 - V q5_0 (6BPW) - FA, 10 - K q6_0 - V iq4_nl (5.5BPW) - FA, 11 - K q5_1 - V q5_0 (5.5BPW) - FA, 12 - K q5_1 - V iq4_nl (5.25BPW) - FA, 13 - K q5_0 - V iq4_nl (5BPW) - FA, 14 - K iq4_nl - V iq4_nl (4.5BPW) - FA, 15 - F16 (16BPW) - FA or not, 16 - K q8_0 - V F16 (12.25BPW) - FA or not, then slower, 17 - K q6_0 - V F16 (11.5BPW) - NO FA, slower, to be tested, 18 - K q5_1 - V F16 (11BPW) - NO FA, slower, 19 - K q5_0 - V F16 (11.75BPW) - NO FA, slower, 20 - K q4_1 - V F16 (10.5BPW) - NO FA, slower, 21 - K q4-0 - V F16 (10.25BPW) - NO FA, slower, 22 - K iq4_nl - V F16 (10.25BPW) - NO FA, slower, to be tested", metavar=('[quantization level 0/1/2/3/4/5/6/7/8/9/10/11/12/13/14/15/16/17/18/19/20/21/22]'), type=check_range(int,0,22), default=0)
+
     advparser.add_argument("--forceversion", help="If the model file format detection fails (e.g. rogue modified model) you can set this to override the detected format (enter desired version, e.g. 401 for GPTNeoX-Type2).",metavar=('[version]'), type=int, default=0)
     advparser.add_argument("--smartcontext", help="Reserving a portion of context to try processing less frequently. Outdated compared to ContextShift, but works with KV Cache quantized unlike ContextShift which doesn't. Not recommended except for the use of KV Cache quantized (KVQ) with FlashAttention (modes 1 to 20).", action='store_true')
     advparser.add_argument("--unpack", help="Extracts the file contents of the Croco.Cpp/KoboldCpp binary into a target directory.", metavar=('destination'), type=str, default="")